MEDICON’23 and CMBEBIH’23: Proceedings of the Mediterranean Conference on Medical and Biological Engineering and Computing (MEDICON) and International ... in Healthcare (IFMBE Proceedings, 93) 3031490614, 9783031490613
124 120 82MB
English Pages [923]
Preface
Contents
Biomedical Signal Processing
Multimodal Registration Algorithm of Plantar Pressure Images and Digital Scanner for the Characterization of the Diabetic Foot in the BASPI/FootLab Laboratory
1 Introduction
2 Methodology
2.1 Protocol
2.2 Pre-Processing
2.3 Multimodal Image Registration of Plantar Pressure Images and Scanner Images
3 Results
4 Discussion
5 Conclusion
References
Semi-Automated Calculation of Baroreflex Sensitivity (BRS) Indices
1 Introduction
2 Methods
2.1 Participants
2.2 Autonomic Nervous System Testing
2.3 Manual Calculation of the Baroreflex Sensitivity Analysis
2.4 Semi-Automated Calculation of the Baroreflex Sensitivity Analysis
2.5 Statistical Analysis
3 Results
4 Discussion
References
Preliminary Assessment of the Samsung Galaxy Watch 5 Accuracy for the Monitoring of Heart Rate and Heart Rate Variability Parameters
1 Introduction
2 Materials and Methods
2.1 Subjects
2.2 Experimental Settings
2.3 Data Processing
2.4 Statistical Analysis
2.5 Results
2.6 Discussion
References
EEG Microstate Clustering to Evaluate Acoustic Stimulation Phase-Locked Targeting of Slow Wave Sleep Activity
1 Introduction
2 Methods
2.1 Dataset
2.2 Data Preprocessing
2.3 Closed-Loop Acoustic Stimulation
2.4 Microstate Analysis
2.5 Statistical Evaluation
3 Results
3.1 Microstate Analysis
4 Discussion
4.1 Limitations
5 Conclusion
References
Development of an Interpretable Model for Improving Differential Diagnosis in Subjects with a Left Ventricular Ejection Fraction Ranging from 40 to 55%
1 Introduction
2 Methods
2.1 Study Population and Protocol
2.2 Heart Rate Variability Acquisition and Processing
2.3 Outliers, Features Selection and Classification
2.4 Nomogram
3 Results
4 Discussion
References
Fractal Characteristics of Retinal Microvascular Network in Alzheimer's Disease and Colon Cancer in Automatically Segmented Fundus Images from the UK Biobank Database
1 Introduction
2 Methods
2.1 Retinal Images
2.2 Participant Selection
2.3 Segmentation of Retinal Vessels, Image Binarization, and Skeletonization
2.4 Box Counting Fractal Analyses and Lacunarity Analyses
2.5 Statistical Analyses
3 Results
3.1 Automatic Segmentation of Blood Vessels
3.2 Age Distribution of Participants
3.3 Box-Counting Dimension and Lacunarity Analyses
4 Discussion
5 Conclusion
References
Classification of Atrial Fibrillation ECG Signals Using 2D CNN
1 Introduction
2 Motivation
3 Implementation
3.1 Preprocessing
3.2 CWT Scalogram of ECG Segment
3.3 CNN Architecture
4 Results
5 Conclusion and Discussion
References
Feature Selection for Arrhythmia Classification Using Statistical Tests
1 Introduction
2 Motivation
3 Implementation
4 Results
5 Conclusion and Discussion
References
Investigating the Physiology Behind Nose Thermal Response to Stress: A Cross-Mapping Approach
1 Introduction
2 Materials and Methods
2.1 Subjects Recruitment and Experimental Set-up
2.2 ECG and Thermal Time Series Processing
2.3 Cross-mapping
2.4 Statistical Analysis
3 Results
4 Discussions and Conclusions
References
Comparing Valence-Arousal and Positive-Negative Affect Models of Affect: A Nonlinear Analysis of Continuously Annotated Emotion Ratings
1 Introduction
2 Materials and Methods
2.1 The Continuously Annotated Signal of Emotion (CASE) Dataset
2.2 The Positive and Negative Affect (PA-NA) Plane
2.3 Phase Space Reconstruction and Fuzzy Entropy Analysis
2.4 Statistical Analysis
3 Results
4 Discussion and Conclusions
References
Postprandial Peak Identification from Continuous Glucose Monitoring Time Series
1 Introduction
2 Participants and Data Collection
3 Algorithm Description
3.1 Algorithm Development
3.2 Algorithm Stages
3.3 Algorithm Assessment
4 Results and Discussion
4.1 Algorithm Performance Under Controlled Conditions (Dataset 1)
4.2 Algorithm Performance Under Free-Living Conditions (Dataset 1)
4.3 Algorithm Performance on the Entire Profiles (Dataset 1)
4.4 Comparison with Other Algorithms Available from Literature
4.5 Algorithm Evaluation on a Different Dataset (Dataset 2)
5 Conclusions
6 Limitations of research
References
Emotional State Evaluation in Driving Simulations: PC Versus Virtual Theater
1 Introduction
2 Methods
2.1 Experimental Protocol
2.2 Signal Analysis
2.3 Statistical Analysis
3 Results
4 Discussion
References
Classification of Physiological States Through Machine Learning Algorithms Applied to Ultra-Short-Term Heart Rate and Pulse Rate Variability Indices on a Single-Feature Basis
1 Introduction
2 Materials and Methods
2.1 Subjects and Experimental Protocol
2.2 Preprocessing and Features Extraction
2.3 Classification Algorithms
3 Results
4 Discussion
4.1 Classification Using Short-Term HRV and PRV Features
4.2 REST vs. HUT Classification Using Ultra-Short-Term Features
5 Conclusion
References
Decomposition of HDsEMG Signals Recorded from a Forearm Extensor Muscle Based on Blind Source Separation
1 Introduction
2 Signal Model and Decomposition Algorithm
3 Signal Processing Procedure and Results
4 Discussion and Conclusion
References
Exploring the Predictability of EEG Signals Timed with the Heartbeat: A Model-Based Approach for the Temporal and Spatial Characterization of the Brain Dynamics
1 Introduction
2 Materials and Methods
2.1 Dataset Description and Pre-processing
2.2 Time-Series Extraction
2.3 Time-Series Characterization
2.4 Statistical Analysis
3 Results
4 Discussion
5 Conclusion
References
Measuring the Balance Between Synergy and Redundancy in Network Systems by Using Information Theory
1 Introduction
2 Methods
2.1 Formulation of the B-Index
2.2 Computation of B-Index
3 Simulation Study
4 Application to the Real Data
4.1 fMRI Data
4.2 In-Vitro Neuronal Culture
5 Conclusion
References
Comparison of Linear Model-Based and Nonlinear Model-Free Directional Coupling Measures: Analysis of Cardiovascular and Cardiorespiratory Interactions at Rest and During Physiological Stress
1 Introduction
2 Material and Methods
2.1 Subjects, Experimental Protocol and Time Series Extraction
2.2 Directional Coupling Measure
2.3 Data and Statistical Analysis
3 Results and Discussion
4 Conclusion
References
Kinematic Characterization of Movements During the Tinetti Test
1 Introduction
2 Materials and Methods
2.1 Subjects and Tasks
2.2 Acquisition and Analysis
3 Results
4 Discussion
5 Conclusions
References
A TMS-EEG Pre-processing Parameters Tuning Study
1 Introduction
2 Methods
2.1 Participants
2.2 Data Acquisition
2.3 Data Processing
2.4 Statistical Analysis
3 Results
4 Discussion
5 Conclusions
References
Non-contact Biopotential Amplifier with Capacitive Driven Right Leg Circuit
1 Introduction and Related Work
2 Electronics System Design
3 Results
4 Conclusion and Further Work
References
Relationship Between Personality and Kinematic Parameters of Handwriting
1 Introduction
2 Materials and Methods
2.1 Subjects and Tasks
2.2 Data Acquisition and Analysis
3 Results
4 Discussion
5 Conclusions
References
Loneliness and Heart Rate in Older Adults
1 Introduction
2 Materials and Methods
2.1 Signal Processing Analysis
3 Results
4 Discussion
References
Filters for Electrocardiogram Signal Processing: A Review
1 Introduction
2 Methods
3 Results
4 Conclusion
References
Medical Physics, Biomedical Imaging and Radiation Protection
Design, Manufacturing and Quality Assessment of 3D-Printed Anthropomorphic Breast Phantom for Mammography
1 Introduction
2 Material and Methods
2.1 Clinical Mammography System
2.2 Characterization of the Material Used in 3D Printing
2.3 Phantom Design
2.4 Testing the Performance of the Phantom
3 Results and Discussion
4 Conclusions
References
Voxelization: Multi-target Optimization for Biomedical Volume Rendering
1 Introduction
2 Methods
2.1 Construction
2.2 Empty Space Skipping
2.3 Maximum Intensity Projection
2.4 Woodcock Tracking
3 Results
4 Discussion and Conclusion
References
Importance of Patient Dose Evaluation and Optimization in Thorax Computed Tomography
1 Introduction
2 Materials and Methods
2.1 Data Collection
2.2 Evaluation of Optimization Practices
3 Results and Discussion
4 Conclusion
References
A Novel X-Ray 3D Histological Method for Paraffinated Prostate Samples
1 Introduction
2 Materials and Methods
2.1 Paraffin Clearance and Eosin Staining
2.2 Micro-CT Setup and Imaging
2.3 Histology
2.4 Analysis
3 Results and Discussion
3.1 Paraffin Clearance and Eosin Staining
3.2 Micro-CT Imaging, Segmentation and Measurements
3.3 Micro-CT vs Histology
4 Conclusions
References
Review of MRI Reporter Genes in Oncology
1 Introduction
2 Methodology
3 Results and Discussion
References
Image Registration Techniques for Independent Acquisitions of Cone Beam Computed Tomography Volumes
1 Introduction
2 Materials and Methods
3 Results and Discussion
4 Conclusion
References
Extremity Bones Segmentation in Cone Beam Computed Tomography, a Novel Approach
1 Introduction
2 Materials and Methods
3 Results
4 Conclusions
References
Deep-Learning Based Automatic Determination of Cardiac Planes in Survey MRI Data
1 Introduction
1.1 Motivation
1.2 State-Of-The-Art Methods
1.3 Presented Work
2 Methodology
2.1 Data Description
2.2 Manual Cardiac Planes Determination
2.3 CNN-Based Cardiac Planes Determination
3 Results and Discussion
4 Conclusion
References
‘3D Printed Breast Phantoms Materials for X-ray Imaging Techniques’
1 Introduction
1.1 Equipment and Materials
1.2 Fused Deposition Modeling Settings (FDM)
1.3 Stereolithography Settings (SLA)
1.4 X-ray Imaging
2 Results
2.1 Comparison of the Linear Theoretical and Experimental Attenuation Coefficients
3 Conclusion
References
Multi-Scale Assessment of Harmonization Efficacy on Resting-State Functional Connectivity
1 Introduction
2 Methods
2.1 Subjects and Experimental Protocol
2.2 MRI Data Pre-Processing
2.3 Seed-Based Functional Connectivity Analysis
2.4 ComBat Harmonization of Functional Connectivity Matrix
2.5 Evaluation of Harmonization Effects on Functional Connectivity Values
2.6 Evaluation of Harmonization Effects on Functional Connectivity Quality Control Metric
3 Results
3.1 ComBat Harmonization Performance on Functional Connectivity
3.2 ComBat Harmonization Effects on Functional Connectivity Quality Control Metric
4 Discussion and Further Conclusions
References
Comparison of Spine Segmentation Algorithms on Clinical Data from Spectral CT of Patients with Multiple Myeloma
1 Introduction
2 Dataset
3 Methods
4 Results and Discussion
5 Conclusion
References
Evaluation of the Effect of Iodine Staining and Sodium Thiosulfate Clearing on the Quality of X-ray Microtomographic Images and Histological Processing of Small Bowel Biopsies
1 Introduction
2 Materials and Methods
2.1 Sample Preparation and ΜCT Imaging
2.2 STS Clearing and Histology
2.3 Evaluation of Images
3 Results
4 Discussion
5 Conclusion
References
Results of Daily Quality Control in Computed Tomography
1 Introduction
2 Materials and Methods
3 Results
4 Discussion
5 Conclusions
References
Design and Implementation of an In-House Built Physical Phantom for Bone Density Measurements
1 Introduction
2 Materials and Methods
2.1 Computational Model
2.2 Printed Model
2.3 Evaluation of the Model
3 Results and Discussion
4 Conclusions
References
Impact of PET and PET/CT Radiotracer for Evaluating Efficacy of CDK4/6 Inhibitor Therapy in HR-Positive HER2-Negative Metastatic Breast Cancer
1 Introduction
2 Methodology
3 Results and Discussion
4 Conclusion
References
Pharmaceutical Engineering
The Effect of the Angiotensin-Converting Enzyme Inhibitor Perindopril on the Lipid Status of Rattus Norvegicus (Berkenhout 1769)
1 Introduction
2 Experimental
3 Results
4 Discussion
5 Conclusion
References
Application of Isotretinoin in the Treatment of Acne Vulgaris: A Questionnaire-Based Survey
1 Introduction
2 Materials and Methods
3 Results
4 Discussion
5 Conclusion
References
Role and Application of Curcumin in Conventional Cancer Therapy
1 Introduction
2 Anticancer Activity of Curcumin
3 Application of Curcumin in Cancer Therapy
4 Conclusion
References
Inclusion Complexation with Randomly Methylated β-Cyclodextrin – An Opportunity to Achieve Greater Solubility of Dimenhydrinate
1 Introduction
2 Materials and Methods
2.1 Preparation of Physical Mixture
2.2 Preparation of Inclusion Complex by Kneading Method
2.3 Phase Solubility Analysis
2.4 Attenuated Total Reflectance - Fourier Transform Infrared Spectroscopy (ATR/FTIR)
3 Results and Discussion
3.1 Phase Solubility Analysis
3.2 ATR/FTIR Analysis
4 Conclusion
References
Cost-Effectiveness of Disinfectant and Antimicrobial Products Usage in Public Transport
1 Introduction
2 Methods
2.1 Adenosine Triphosphate (ATP) Measurements
2.2 Pharmacoeconomic Analysis
3 Results
4 Discussion
5 Conclusion
References
Association Between Serum Free Fatty Acids and Total Bilirubin Concentrations in Bosnian Individuals with Diabetes Mellitus Type 2
1 Introduction
2 Materials and Methods
3 Results and Discussion
4 Conclusion
References
HPLC-UV Determination and Comparison of Extracted Corticosteroids Content with Two Methods
1 Introduction
2 Materials and Methods
2.1 First Method – Operating Conditions
2.2 Second Method – Operating Procedures
3 Results and Discussion
3.1 Results of the First Method
3.2 Results of the Second Method
4 Conclusion
References
Can Microneedles Revolutionize Ocular Drug Delivery?
1 Introduction
2 Methods
3 Microneedles
4 Microneedle Systems as Ocular Drug Delivery Systems
5 Conclusion
References
Analysis of Lysozyme as Biomarker in Saliva
1 Introduction
2 Materials and Methods
2.1 Sample Collection and Storage
2.2 Chemicals
2.3 Turbidimetric Method for Lysozyme Activity Analysis
2.4 Ethics Statement
2.5 Statistical Analysis
3 Results
4 Discussion
5 Conclusion
References
Characterization of the Chemical Substance Niacinamide
1 Introduction
2 Materials and Methods
2.1 Chemicals
2.2 Identification According to Ph. Yug. IV
2.3 Identification and Determination According to Ph. Eur. 9
2.4 Preparation of Solutions for the Characterization of Chemical Substance
2.5 Preparation of Solutions for Determining the Linearity of the UV-VIS Method
3 Results and Discussion
3.1 Identification According to Ph. Yug. IV
3.2 Identification and Determination According to Ph. Eur. 9
3.3 Characterization of a Chemical Substance
3.4 Linearity of the UV-VIS Method for the Determination of Niacinamide
4 Conclusion
References
In Vitro Aerodynamic Comparison of Protective Masks
1 Introduction
2 Materials and Methods
2.1 Chemicals, Apparatus and Samples
2.2 Principle and Procedure
3 Results and Discussion
3.1 Reference Standards
3.2 Positive Control
3.3 Efficiency of Filtration by Cascade Impactor Sections and by Particles in the Size Range of Interest
3.4 Method Selection
4 Conclusion
References
Curl Cream Development – A Pilot Study
1 Introduction
1.1 The Characteristics of Curly Hair
2 Materials and Methods
2.1 Materials
2.2 Methods
3 Results and Discussion
4 Conclusion
References
The Total Phenols and Total Flavonoids Content of Petioles Sweet Cherry (Prunus Avium L.)
1 Introduction
2 Material and Methods
2.1 Sampling
2.2 Reagents and Chemicals
2.3 Preparation of Extracts
2.4 Spectrophotometric Determination of Total Phenol Content (TPC) and Total Flavonoid Content (TFC)
3 Results and Discussion
3.1 Mass Fractions of Extracts of Dry Cherry Petioles
3.2 Total Phenolic Content (TPC)
3.3 Total Flavonoid Content (TFC)
4 Conclusion
References
Determination of Capsaicin and Dihydrocapsaicin Content by HPLC Method in Products Purchased Online
1 Introduction
2 Materials and Methods
2.1 Chemicals
2.2 Absorption Spectrum of Capsaicin and Dyhidrocapsaicin
2.3 High Performance Liqiud Chromatography (HPLC)
2.4 Linearity and Repeatability
2.5 Sample Preparation
3 Results and Discussion
3.1 Absorption Spectrum of Capsaicin and Dihydrocapsaicin
3.2 High Performance Liqiud Chromatography (HPLC)
3.3 Linearity and Repeatability
3.4 Determination of Capsaicin and Dihydrocapsaicin in Samples
4 Conclusions
References
Artificial Intelligence and Machine Learning
Explainable CNN-Based Cardiac Amyloidosis Classification from PET Images Through Manifold Learning
1 Introduction
2 Materials and Methods
3 Results
4 Conclusion
References
Evaluation of Bacterial Biofilm Category Change Due to the Use of Different Signaling Molecules Using Random Forest Classifier
1 Introduction
2 Related Work
3 Random Forest Classifier
4 Dataset Analysis
5 Methods
5.1 Combining Different Inputs for Each Classification
5.2 Model Development
5.3 Key Classification Metrics
5.4 Cross-Validation
6 Results and Discussion
7 Conclusion
References
Evaluating Bilateral Surface EMG Features for Automatic Identification of Gait Phase Transitions in Ground Walking Conditions
1 Introduction
2 Materials and Methods
2.1 Participants and Experimental Procedure
2.2 Signal Processing
2.3 Data Preparation and Training Procedure
2.4 Detection of Transition Events
3 Results
4 Discussion and Conclusions
References
A Vital Signs-Driven Approach for Clustering the Responses to Diuretics Treatment in Premature Newborns
1 Introduction
2 Materials and Methods
3 Results and Discussion
4 Conclusions
References
Identification of Atypical Cardiac Patterns Before and After Exercising Using Artificial Intelligence and Eulerian Video Magnification
1 Introduction
2 Methodology
2.1 Eulerian video magnification
2.2 Identification of Cardiac Atypical Patterns
2.3 Graphic Interface
3 Results
3.1 Eulerian Video Magnification
3.2 Detection of Atypical Cardiac Patterns
4 Discussion
5 Conclusions
References
Clinical Natural Language Processing and Health Interoperability to Support Knowledge Management and Governance in Rare Cancers
1 Introduction
2 Methodology
3 Results
3.1 Data Governance and Sovereignty
3.2 Federated Learning and Artificial Intelligence Approaches
3.3 Multimodal Data Navigator
3.4 Ecosystem Validation
4 Conclusions
References
Deep Learning Enabled Acute Ischemic Stroke Lesion Segmentation for Smart Healthcare Systems
1 Introduction
1.1 Motivation and Contributions
2 Related Literature
3 Materials and Methods
3.1 Dataset
3.2 Experimental Setting
3.3 Preprocessing
3.4 Tissue Mask Segmentation
3.5 Smoothing
3.6 Our Model
3.7 Abnormal Hypoattenuation Detection
3.8 Assessing Heatmap Impact on Quality of Expert Annotations
4 Experimental Results
4.1 Hypothesis-1: Relation of Gender and Type of Stroke
4.2 Hypothesis-2: Increase in Number of Cases Can Affect Model Performance in Specific Fields
5 Discussion
6 Conclusion
References
When an Explanation is not Enough: An Overview of Evaluation Metrics of Explainable AI Systems in the Healthcare Domain
1 Introduction
2 Methods
3 Metrics for Evaluating Explainable AI
3.1 Domain Independent Metrics with Human in the Loop
3.2 Domain Independent Metrics Without Human in the Loop
3.3 Healthcare Domain Metrics
4 Discussion
5 Conclusion
References
Liver Diseases Classification Using Machine Learning Algorithms
1 Introduction
2 Materials and Methods
2.1 Dataset and Data Preparation
2.2 Tools, Methods and Model Building
3 Results
4 Conclusion
References
Interpretability and Personalization Aspects in the Development of Clinical Risk Assessment Models
1 Introduction
2 Methods
2.1 GRACE Risk Score: Clinical Reference
2.2 Interpretable Models
2.3 Global Evaluation
3 Results
3.1 Dataset
3.2 Data Preprocessing
3.3 Preliminary Analysis
3.4 GRACE Model Evaluation
3.5 ML Models Evaluation
4 Discussion
5 Conclusions
References
Ten Year Cardiovascular Risk Estimation: A Machine Learning Approach
1 Introduction
2 Materials and Methods
2.1 Tools
2.2 Dataset
2.3 Experiment
3 Results
4 Conclusion
References
Thyroid Hormones Parameter-Based Classification of Patient Health Status: An Analysis of Machine Learning Techniques
1 Introduction
2 Materials and Methods
2.1 Tools
2.2 Dataset
2.3 Machine Learning Algorithms
2.4 Model Evaluation Tools and Metrics
3 Results and Discussion
4 Conclusion
5 Future Work
References
Artificial Intelligence-Based Ultrasound Imaging Classification for Infant Neurological Impairment Disorders: A Review
1 Introduction
2 Methods
3 Results
3.1 Proposed Methodology for Fetal Neurological Risk Assessment Using AI
4 Conclusion
References
Cardiovascular Health in AI: A Comprehensive Overview to Acute Myocardial Infarction Prediction
1 Introduction
2 Literature Review
3 Materials and Methods
3.1 Methodology
3.2 Acute Myocardial Infarction Database
4 Results and Discussion
5 Conclusion
References
Sepsis Detection Using Features Extracted from Photoplethysmography
1 Introduction
2 Material and Methods
2.1 Dataset
2.2 Feature Extraction
2.3 XGBoost Classifier Training
3 Results
3.1 XGBoost Results
3.2 Previous Findings
4 Discussion
References
Digital Histopathological Discrimination of Label-Free Healthy Tissues by Decision Tree Classifier
1 Introduction
2 Materials and Methods
2.1 Phase Contrast Microscopy
2.2 Phase Contrast Parameters
3 Results and Discussion
4 Conclusions
References
Heart Disease Prediction Using Logistic Regression Machine Learning Model
1 Introduction
1.1 Heart Diseases
1.2 Machine Learning Use in Medical Field
2 Methods
2.1 Programming Language and Libraries Used
2.2 Data Used
2.3 Model for Predicting Risk of a Patient Having Heart Disease
3 Results
3.1 Correlation of the Data in Dataset
4 Conclusion
References
Class Probability Distributions of a Neural Network Classifier of Multiple Sclerosis Lesions on Quantitative Susceptibility Mapping
1 Introduction
2 Materials and Methods
3 Results
4 Conclusion
References
Forecasting Icterus with Machine Learning: An Advanced Classification Model Analysis
1 Introduction
2 Material and Methods
2.1 Tools
2.2 Dataset
2.3 Models
3 Results and Discussion
3.1 Logistic Regression
4 Conclusion
References
Evaluation of Deep Learning Techniques for Automatic Lesion Segmentation in Mammography Images
1 Introduction
2 Dataset
3 Methodology
3.1 Pre-Processing
3.2 Models
3.3 Evaluation Metrics
4 Results
5 Discussion
6 Conclusion
References
Integrating Machine Learning in Clinical Decision Support for Heart Failure Diagnosis: Case Study
1 Introduction
2 Methods
3 Results and Discussion
4 Conclusion
References
The Security Framework: Determined and AI Impact?
1 Introduction
2 Technology as the Security’s Negative Variable
2.1 Cyber Challenges?
3 Artificial Intelligence and the Economic Race
4 Conclusion
References
The Future of (Individual) Security—Vision 2.0?
1 Introduction
2 Individual’s Security Practical Implications: e-Health Example
3 Vision of the Future of Security?
3.1 Cyber Challenges?
4 New Paradigm for the New Concept of Security?
5 Conclusion
References
Forecasting Meningitis with Machine Learning: An Advanced Classification Model Analysis
1 Introduction
2 Materials and Methods
2.1 Dataset
2.2 Tools
3 Data Analysis
4 Results
4.1 Model Evaluation Metrics
4.2 The Hyper Parameter Models
5 Conclusion
References
Analysis of Predictive Parameters in Prediction of the Occurrence of Type 2 Diabetes Using Machine Learning Algorithms
1 Introduction
2 Methods and Dataset
2.1 Dataset
2.2 Tools and Methods
3 Results
4 Discussion
5 Conclusion
References
Health Informatics, e-Health and Telemedicine
Challenges and Outlook to Designing Cutting-Edge Mixed Reality Technologies of Human Pose Estimation
1 Introduction
2 State-of-Art of Human Tracking Techniques
3 Review: Studies Selection
3.1 OpenPose
3.2 PoseNet
3.3 MoveNet
3.4 MediaPipe
3.5 Projects
3.6 Proprietary Algorithms
4 Discussions
5 Conclusions
References
Potentially Inappropriate Medications in Geriatric Patients with Type 2 Diabetes Mellitus: Practical Software Solution for Healthcare Professionals
1 Introduction
2 Materials and Methods
3 Expected Results
4 Conclusion
References
Addressing Rapidly Aging Society Challenges Through Health 4.0
1 Introduction
2 Health 4.0
3 Health 4.0 Technologies
4 Discussion
5 Conclusions
References
A Virtual Reality-Based Setting to Investigate How Environments and Emotionally-Laden Stimuli Interact and Compete for Accessing Consciousness
1 Introduction
2 Methods
2.1 Experimental Sample
2.2 Experimental Setting
2.3 Virtual Scenarios
2.4 Experimental Procedure
2.5 Data Analysis
3 Results
4 Discussion
References
Smart Sensors for Daily-Life Data Collection Toward Precision and Personalized Medicine: The TOLIFE Project Approach
1 Introduction
2 Methods
2.1 Smart Sensors Definition and Architecture
2.2 Smartwatch
2.3 Smartphone
2.4 Smartshoes
2.5 Smart Mattress Cover and Environmental Unit
2.6 Mini-spirometer
3 Data Integration and Communication
4 Discussion and Conclusion
References
Exploring the Progress of Digital Transformation in the Healthcare System of Bosnia and Herzegovina
1 Introduction
2 Current State and Strategic Orientation Toward Digital Transformation of Healthcare in Bosnia and Herzegovina
3 Progress of Digital Transformation in the Healthcare System of Bosnia and Herzegovina
4 Conclusion
References
Medical Engineering, Health Informatics and Medical Imaging (ME&HI) Profession and Curriculum
1 Introduction
1.1 Term-Phrases for “Medical Instruments and Devices” and “Health Information Systems”
1.2 Included and Excluded Term-Phrases for the First Two Pillars (ME&HI)
1.3 Term-Phrases for “Medical Image Acquisition & Management, Medical Physics and Diagnostic Radiology” Pillar
1.4 Rationale for Choosing the Name ME&HI (First 2 Pillars) and MI (3rd Pillar)
1.5 Overal Naming of Study (All – 3 – Pillars)
2 Usage of ME&HI in Science, Professions, Education and Training
2.1 ME&HI Usage in Journals and Conferences
2.2 ME&HI Usage in Universities
2.3 ME&HI Usage in Scientific and Professional Organizations
2.4 ME&HI Usage in Training Programs
3 Who Receives ME&HI Knowledge and Where/how They Use It?
4 Examples of Core Knowledge and Complementary Skills of ME&HI
4.1 ME&HI Complementary Skills
4.2 ME&HI Core Knowledge for Intensive Care Units (ICUs)
5 Discussion
6 Conclusions and Future Activities
References
Patients’ Health Information Exchange Within the European Union: A Croatian Case Study
1 Introduction
2 e-Health Digital Service Infrastructure
3 Croatian e-Health: The Outlook
3.1 Croatian Legislation Overview
3.2 Strategic Framework for e-Health Development
4 Discussion
4.1 COVID-19 Pandemic Experience
4.2 Towards the European Health Data Space
4.3 Ethical Considerations
5 Conclusion
References
Interaction Between Pharmaceutical Companies and the Public During the COVID-19 Pandemic—A Twitter Analysis
1 Introduction
2 Methods
3 Results
4 Discussion
5 Conclusions
References
Innovations in the Food System—Towards Next Generation of Multidisciplinary Initiatives
The Many Faces of E. Faecium: From Probiotics to Pathogenesis
1 Introduction
2 Hospital-Associated E. Faecium and the Problem with VRE
3 Community-Associated E. faecium and E. lactis in Food Products
4 Conclusion
References
Antimicrobial Resistance in the Food Chain—Are We at the Point Where There is no Time to Wait?
1 Antimicrobial Resistance and Its Impact to Public Health
1.1 Antimicrobials and Resistance
1.2 Mechanisms of AMR
2 Antimicrobial Resistance in the Food Supply Chain
3 Prevention Measures to Combat Resistance—A Farm to Fork Approach
3.1 A Holistic Approach to Combat Antimicrobial Resistance
4 Conclusion
References
HPLC Q-TOF LC/MS Analysis of Inulin in Foods: Development of an Innovative Chromatography Method for Nutritional Enhancement
1 Introduction
2 Methods and Materials
3 Results
4 Discussion
5 Conclusion
References
Whole Genome Sequencing for Food Safety, Clinical and Public Health Microbiology
1 Introduction
2 Whole Genome Sequencing in Food Microbiology
3 WGS in Medical and Public Health Microbiology
4 Conclusion
References
Green Chemistry
Analysis of Atmospheric Concentrations of PM 10 and PM 2.5 and Their Impact on the Health Status of the Population in Urban Industrial Cities of Bosnia and Herzegovina
1 Introduction
2 Methodology
2.1 Data
3 Results and Discussion
4 Conclusion
References
Optimization of Glyphosate Adsorption Conditions on Pyrophyllite
1 Introduction
2 Materials and Methods
2.1 Pyrophyllite
2.2 Chemicals
2.3 Instrumentation
2.4 Quantification of Glyphosate
2.5 Determination of Point Zero Charge (pHpzc) of Pyrophyllite
2.6 Preparation of Samples
2.7 Removal of Glyphosate
2.8 Adsorption Isotherms
2.9 Kinetic Studies
3 Results and Discussion
3.1 PHpzc of Pyrophyllite
3.2 Influence of Adsorbent Dose
3.3 Influence of pH
3.4 Influence of Contact Time
3.5 Influence of Temperature
3.6 Adsorption Isotherms
3.7 Adsorption Kinetics
4 Conclusion
References
Understanding Main Drivers of Global Decarbonization
1 Introduction
2 Restructuring of Energy System in XXI Century
3 Global Imbalances in Energy Demand and Supply—Neglected, But Critical
4 Main Drivers of Decarbonization
4.1 Market Price Signals and CO2 Pricing
4.2 Renewable Energy Sources (RES)
4.3 Diversification of Energy Supply—Usage of Alternative Technologies
5 Conclusion
References
Author Index
Recommend Papers
![Data Analysis and Optimization for Engineering and Computing Problems: Proceedings of the 3rd EAI International Conference on Computer Science and Engineering and Health Services [1st ed.]
9783030481483, 9783030481490](https://ebin.pub/img/200x200/data-analysis-and-optimization-for-engineering-and-computing-problems-proceedings-of-the-3rd-eai-international-conference-on-computer-science-and-engineering-and-health-services-1st-ed-9783030481483-9783030481490.jpg)
- Author / Uploaded
- Almir Badnjevic
- Lejla Gurbeta Pokvić
File loading please wait...
Citation preview
IFMBE Proceedings 93
Almir Badnjević Lejla Gurbeta Pokvić Editors
MEDICON’23 and CMBEBIH’23 Proceedings of the Mediterranean Conference on Medical and Biological Engineering and Computing (MEDICON) and International Conference on Medical and Biological Engineering (CMBEBIH), September 14–16, 2023, Sarajevo, Bosnia and Herzegovina—Volume 1: Imaging, Engineering and Artificial Intelligence in Healthcare
IFMBE Proceedings Series Editor Ratko Magjarevi´c, Faculty of Electrical Engineering and Computing, ZESOI, University of Zagreb, Zagreb, Croatia
Associate Editors Piotr Łady˙zy´nski, Warsaw, Poland Fatimah Ibrahim, Department of Biomedical Engineering, Faculty of Engineering, Universiti Malaya, Kuala Lumpur, Malaysia Igor Lackovic, Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia Emilio Sacristan Rock, Mexico City, Mexico
93
The IFMBE Proceedings Book Series is an official publication of the International Federation for Medical and Biological Engineering (IFMBE). The series gathers the proceedings of various international conferences, which are either organized or endorsed by the Federation. Books published in this series report on cutting-edge findings and provide an informative survey on the most challenging topics and advances in the fields of medicine, biology, clinical engineering, and biophysics. The series aims at disseminating high quality scientific information, encouraging both basic and applied research, and promoting world-wide collaboration between researchers and practitioners in the field of Medical and Biological Engineering. Topics include, but are not limited to: • • • • • • • • • •
Diagnostic Imaging, Image Processing, Biomedical Signal Processing Modeling and Simulation, Biomechanics Biomaterials, Cellular and Tissue Engineering Information and Communication in Medicine, Telemedicine and e-Health Instrumentation and Clinical Engineering Surgery, Minimal Invasive Interventions, Endoscopy and Image Guided Therapy Audiology, Ophthalmology, Emergency and Dental Medicine Applications Radiology, Radiation Oncology and Biological Effects of Radiation Drug Delivery and Pharmaceutical Engineering Neuroengineering, and Artificial Intelligence in Healthcare
IFMBE proceedings are indexed by SCOPUS, EI Compendex, Japanese Science and Technology Agency (JST), SCImago. They are also submitted for consideration by WoS. Proposals can be submitted by contacting the Springer responsible editor shown on the series webpage (see “Contacts”), or by getting in touch with the series editor Ratko Magjarevic.
Almir Badnjevi´c · Lejla Gurbeta Pokvi´c Editors
MEDICON’23 and CMBEBIH’23 Proceedings of the Mediterranean Conference on Medical and Biological Engineering and Computing (MEDICON) and International Conference on Medical and Biological Engineering (CMBEBIH), September 14–16, 2023 Sarajevo, Bosnia and Herzegovina—Volume 1: Imaging, Engineering and Artificial Intelligence in Healthcare
Editors Almir Badnjevi´c Verlab Research Institute for Biomedical Engineering Medical Devices, and Artificial Intelligence Sarajevo, Bosnia and Herzegovina
Lejla Gurbeta Pokvi´c Verlab Research Institute for Biomedical Engineering Medical Devices, and Artificial Intelligence Sarajevo, Bosnia and Herzegovina
ISSN 1680-0737 ISSN 1433-9277 (electronic) IFMBE Proceedings ISBN 978-3-031-49061-3 ISBN 978-3-031-49062-0 (eBook) https://doi.org/10.1007/978-3-031-49062-0 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland Paper in this product is recyclable.
Preface
Dear colleagues, It is with immense gratitude and delight that we look back on the successful joint event of the 16th Mediterranean Conference on Medical and Biological Engineering and Computing (MEDICON) and the 5th International Conference on Medical and Biological Engineering (CMBEBIH), which took place in the city of Sarajevo, Bosnia and Herzegovina, from September 14th to 16th, 2023. During those three unforgettable days, we had the privilege of witnessing engaging keynote speeches by esteemed experts in the field. We were thrilled to see attendees actively participating in scientific and industry sessions, passionately contributing to interactive panel discussions, and connecting with peers from diverse backgrounds and expertise. This conference served as an exceptional platform, uniting researchers, scientists, engineers, professionals, and public sector representatives from various disciplines within the medical and biological engineering field. The meticulously curated program of the conference encompassed a wide range of captivating topics, focusing on the transformative power of digitalization in healthcare and the environment. The introduction of digital innovation hubs proved to be a resounding success, enabling interactive demonstrations and fostering collaboration among attendees. Additionally, the “Meet the Editor” session provided valuable insights into publishing research, leaving a lasting impact on the attendees. We are truly humbled by the active participation and enthusiastic engagement of all those who attended, making this conference a resounding success. The networking opportunities that were availed allowed for fruitful connections, idea exchanges, and the forging of new collaborations that have the potential to shape the future of medical and biological engineering. As we reflect on the conference, we are filled with profound gratitude for the shared experiences, the sparks of curiosity ignited, and the passion for innovation that was palpable throughout the event. It is our sincere hope that the knowledge gained and the connections made will continue to foster progress in the field of medical and biological engineering. We hope that this book, which compiles the valuable contributions and insights presented during the conference, will be a valuable addition to the advancement of biomedical engineering. Through sharing these ideas and research findings, we aspire to further propel the field’s growth and pave the way for future innovations and breakthroughs.
vi
Preface
Once again, we express our heartfelt thanks to all participants, speakers, sponsors, and the organizing team for making this conference an unforgettable experience. We are honored to have had each of you contribute to its success. Prof. Dr. Almir Badnjevi´c Prof. Dr. Lejla Gurbeta Pokvi´c Conference Chairs
Contents
Biomedical Signal Processing Multimodal Registration Algorithm of Plantar Pressure Images and Digital Scanner for the Characterization of the Diabetic Foot in the BASPI/FootLab Laboratory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yeison Luna and Martha Zequera Semi-Automated Calculation of Baroreflex Sensitivity (BRS) Indices . . . . . . . . . Magdalena Krbot Skori´c, Ivan Adamec, Ivan Moštak, Nika Višnji´c, Mario Cifrek, and Mario Habek Preliminary Assessment of the Samsung Galaxy Watch 5 Accuracy for the Monitoring of Heart Rate and Heart Rate Variability Parameters . . . . . . . Gianluca Rho, Francesco Di Rienzo, Carlotta Marinai, Francesca Giannetti, Lucia Arcarisi, Pasquale Bufano, Michele Zanoletti, Francesca Righetti, Carlo Vallati, Marco Laurino, Nicola Carbonaro, Alessandro Tognetti, and Alberto Greco EEG Microstate Clustering to Evaluate Acoustic Stimulation Phase-Locked Targeting of Slow Wave Sleep Activity . . . . . . . . . . . . . . . . . . . . . . Filip Cerny, Vaclava Piorecka, Jan Strobl, Daniela Dudysova, Jana Koprivova, and Marek Piorecky Development of an Interpretable Model for Improving Differential Diagnosis in Subjects with a Left Ventricular Ejection Fraction Ranging from 40 to 55% . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Katerina Iscra, Miloš Ajˇcevi´c, Aleksandar Miladinovi´c, Laura Munaretto, Jacopo Giulio Rizzi, Marco Merlo, and Accardo Agostino Fractal Characteristics of Retinal Microvascular Network in Alzheimer’s Disease and Colon Cancer in Automatically Segmented Fundus Images from the UK Biobank Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Isidora Rubeži´c, Miroslav Radunovi´c, Dejan Babi´c, Tomo Popovi´c, and Nataša Popovi´c Classification of Atrial Fibrillation ECG Signals Using 2D CNN . . . . . . . . . . . . . Amina Tihak, Lejla Smajlovic, and Dusanka Boskovic
3
13
22
31
41
49
57
viii
Contents
Feature Selection for Arrhythmia Classification Using Statistical Tests . . . . . . . . Amina Tihak, Amna Grahic, and Dusanka Boskovic Investigating the Physiology Behind Nose Thermal Response to Stress: A Cross-Mapping Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Federica Gioia, Mimma Nardelli, Enzo Pasquale Scilingo, and Alberto Greco Comparing Valence-Arousal and Positive-Negative Affect Models of Affect: A Nonlinear Analysis of Continuously Annotated Emotion Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Andrea Gargano, Enzo Pasquale Scilingo, and Mimma Nardelli Postprandial Peak Identification from Continuous Glucose Monitoring Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Aikaterini Archavli, Harpal Randeva, and Natasha Khovanova
66
77
86
96
Emotional State Evaluation in Driving Simulations: PC Versus Virtual Theater . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 Rita Laureanti, Simone Mentasti, Alessandro Gabrielli, Matteo Matteucci, and Luca Mainardi Classification of Physiological States Through Machine Learning Algorithms Applied to Ultra-Short-Term Heart Rate and Pulse Rate Variability Indices on a Single-Feature Basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 Marta Iovino, Ivan Lazic, Tatjana Loncar-Turukalo, Michal Javorka, Riccardo Pernice, and Luca Faes Decomposition of HDsEMG Signals Recorded from a Forearm Extensor Muscle Based on Blind Source Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 Šimun Krmek, Mario Cifrek, Yueming Gao, and Željka Luˇcev Vasi´c Exploring the Predictability of EEG Signals Timed with the Heartbeat: A Model-Based Approach for the Temporal and Spatial Characterization of the Brain Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 Valeria Rosalia Vergara, Chiara Bara, Riccardo Pernice, Andrea Zaccaro, Francesca Ferri, Luca Faes, and Yuri Antonacci Measuring the Balance Between Synergy and Redundancy in Network Systems by Using Information Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 Yuri Antonacci, Gorana Mijatovic, Laura Sparacino, Simone Valenti, Gianvincenzo Sparacia, Daniele Marinazzo, Sebastiano Stramaglia, and Luca Faes
Contents
ix
Comparison of Linear Model-Based and Nonlinear Model-Free Directional Coupling Measures: Analysis of Cardiovascular and Cardiorespiratory Interactions at Rest and During Physiological Stress . . . . . . . . . . . . . . . . . . . . . . . . 155 Chiara Barà, Riccardo Pernice, Laura Sparacino, Yuri Antonacci, Michal Javorka, and Luca Faes Kinematic Characterization of Movements During the Tinetti Test . . . . . . . . . . . . 164 Alessandra Raffini, Francesco Bassi, Miloš Ajˇcevi´c, Aleksandar Miladinovi´c, and Agostino Accardo A TMS-EEG Pre-processing Parameters Tuning Study . . . . . . . . . . . . . . . . . . . . . . 172 Elena Bondi, Viviana Pescuma, Yara Massalha, Marta Pizzolante, Alice Chirico, Giandomenico Schiena, Anna Maria Bianchi, Paolo Brambilla, and Eleonora Maggioni Non-contact Biopotential Amplifier with Capacitive Driven Right Leg Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 Dino Cindri´c, Luka Klai´c, Antonio Staneši´c, and Mario Cifrek Relationship Between Personality and Kinematic Parameters of Handwriting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 Francesco Bassi, Alessandra Raffini, Miloš Ajˇcevi´c, Aleksandar Miladinovi´c, Lisa Di Blas, and Agostino Accardo Loneliness and Heart Rate in Older Adults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 Raquel Cervigón, Samuel Ruipérez-Campillo, José Millet, and Francisco Castells Filters for Electrocardiogram Signal Processing: A Review . . . . . . . . . . . . . . . . . . 204 Elma Kandi´c and Lejla Gurbeta Pokvi´c Medical Physics, Biomedical Imaging and Radiation Protection Design, Manufacturing and Quality Assessment of 3D-Printed Anthropomorphic Breast Phantom for Mammography . . . . . . . . . . . . . . . . . . . . . . 221 Elma Huselji´c, Senad Odžak, Adnan Beganovi´c, Almasa Odžak, Adi Pandži´c, and Merim Jusufbegovi´c Voxelization: Multi-target Optimization for Biomedical Volume Rendering . . . . 232 Elena Denisova, Leonardo Manetti, Leonardo Bocchi, and Ernesto Iadanza
x
Contents
Importance of Patient Dose Evaluation and Optimization in Thorax Computed Tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242 ˇ Belkisa Hani´c, Lejla M. Civa, Mustafa Busuladži´c, Azra Gazibegovi´c-Busuladži´c, Amra Skopljak-Beganovi´c, and Adnan Beganovi´c A Novel X-Ray 3D Histological Method for Paraffinated Prostate Samples . . . . 252 Santiago Laguna-Castro, Teemu Tolonen, Brian Mphande, Jari Hyttinen, and Antti Kaipia Review of MRI Reporter Genes in Oncology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263 ˇ Adna Softi´c, Ivana Ceko, Zerina Kali´c, Nejla Piri´c, Emina Mrdanovi´ c, and Elma Imamovi´c Image Registration Techniques for Independent Acquisitions of Cone Beam Computed Tomography Volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270 Diletta Pennati, Leonardo Manetti, Ernesto Iadanza, and Leonardo Bocchi Extremity Bones Segmentation in Cone Beam Computed Tomography, a Novel Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278 Eleonora Tiribilli, Leonardo Manetti, Leonardo Bocchi, and Ernesto Iadanza Deep-Learning Based Automatic Determination of Cardiac Planes in Survey MRI Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 Jan Jurca, Vratislav Harabis, Roman Jakubicek, Tomas Holecek, Petra Nemcekova, Petr Ourednicek, and Jiri Chmelik ‘3D Printed Breast Phantoms Materials for X-ray Imaging Techniques’ . . . . . . . 293 Aris Dermitzakis, Martin Pichotka, Antzela Petrai, Moritz Weigt, and Nicolas Pallikarakis Multi-Scale Assessment of Harmonization Efficacy on Resting-State Functional Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301 Emma Tassi, Federica Goffi, Maria Gloria Rossetti, Marcella Bellani, Benedetta Vai, Federico Calesella, Francesco Benedetti, Anna Maria Bianchi, Paolo Brambilla, and Eleonora Maggioni Comparison of Spine Segmentation Algorithms on Clinical Data from Spectral CT of Patients with Multiple Myeloma . . . . . . . . . . . . . . . . . . . . . . . 309 Michal Nohel, Roman Jakubicek, Lenka Blazkova, Vlastimil Valek, Marek Dostal, Petr Ourednicek, and Jiri Chmelik
Contents
xi
Evaluation of the Effect of Iodine Staining and Sodium Thiosulfate Clearing on the Quality of X-ray Microtomographic Images and Histological Processing of Small Bowel Biopsies . . . . . . . . . . . . . . . . . . . . . . . 318 Aino Reunamo, Markus Hannula, Katri Lindfors, Teemu Tolonen, Kalle Kurppa, and Jari Hyttinen Results of Daily Quality Control in Computed Tomography . . . . . . . . . . . . . . . . . 325 ˇ Hatina Corbi´ c, Adnan Beganovi´c, Mahira Redži´c, Adnan Šehi´c, Nusret Salkica, and Jasmina Bajrovi´c Design and Implementation of an In-House Built Physical Phantom for Bone Density Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338 Nikolay Dukov, Kristina Bliznakova, Iliyan Kolev, Yanka Baneva, Georgi Valchev, and Zhivko Bliznakov Impact of PET and PET/CT Radiotracer for Evaluating Efficacy of CDK4/6 Inhibitor Therapy in HR-Positive HER2-Negative Metastatic Breast Cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 Elma Imamovi´c, Anisa Veledar, Zerina Kali´c, Adna Softi´c, Emina Mrdanovi´ c, and Merima Smajlhodži´c-Deljo Pharmaceutical Engineering The Effect of the Angiotensin-Converting Enzyme Inhibitor Perindopril on the Lipid Status of Rattus Norvegicus (Berkenhout 1769) . . . . . . . . . . . . . . . . . 353 Elma Haskovi´c, Azra Doli´canin, Edhem Haskovi´c, Safija Herenda, and Orhan Lepara Application of Isotretinoin in the Treatment of Acne Vulgaris: A Questionnaire-Based Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 Amila Ahmetaševi´c-Demirovi´c, Dina Lagumdžija, and Belma Pehlivanovi´c Kelle Role and Application of Curcumin in Conventional Cancer Therapy . . . . . . . . . . 377 Belma Pehlivanovi´c Kelle, Lejla Kurtali´c, Aida Šapˇcanin, and Fahir Beˇci´c Inclusion Complexation with Randomly Methylated β-Cyclodextrin – An Opportunity to Achieve Greater Solubility of Dimenhydrinate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383 Lamija Hindija, Jasmina Hadžiabdi´c, Ognjenka Rahi´c, Amina Tucak-Smaji´c, Merima Šahinovi´c, and Edina Vrani´c
xii
Contents
Cost-Effectiveness of Disinfectant and Antimicrobial Products Usage in Public Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392 Emina Mrdanovi´ c, Lejla Osmanbegovi´c, Merima Smajhodži´c-Deljo, Adna Softi´c, Naida Babi´c-Jordamovi´c, and Haris Vrani´c Association Between Serum Free Fatty Acids and Total Bilirubin Concentrations in Bosnian Individuals with Diabetes Mellitus Type 2 . . . . . . . . . 402 Ša´cira Mandal HPLC-UV Determination and Comparison of Extracted Corticosteroids Content with Two Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412 M. Daci´c, Alija Uzunovi´c, Larisa Alagi´c-Džambi´c, and Saša Pilipovi´c Can Microneedles Revolutionize Ocular Drug Delivery? . . . . . . . . . . . . . . . . . . . . 425 Merima Šahinovi´c, Amina Tucak-Smaji´c, Kenan Muhamedagi´c, Lamija Hindija, Ognjenka Rahi´c, Jasmina Hadžiabdi´c, Edina Vrani´c, ˇ c and Ahmet Ceki´ Analysis of Lysozyme as Biomarker in Saliva . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434 Emina Haskovi´c, Lejda Uzunovi´c, Tanja Duji´c, Aziz Šukalo, Meliha Mehi´c, Maja Malenica, Tamer Bego, Neven Meseldži´c, Selma Imamovi´c Kadri´c, and Una Glamoˇclija Characterization of the Chemical Substance Niacinamide . . . . . . . . . . . . . . . . . . . 443 M. Emira, M. Daci´c, and Alija Uzunovi´c In Vitro Aerodynamic Comparison of Protective Masks . . . . . . . . . . . . . . . . . . . . . 452 ˇ car A. Uzunovi´c, E. Mlivo, M. Daci´c, S. Pilipovi´c, M. Šupuk, and H. Canˇ Curl Cream Development – A Pilot Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463 Hena Brki´c, Emina Sarajli´c, and Alisa Elezovi´c The Total Phenols and Total Flavonoids Content of Petioles Sweet Cherry (Prunus Avium L.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469 Mirsada Salihovi´c, Mirha Pazalja, Ajla Špago, Elma Veljovi´c, and Selma Špirtovi´c-Hali-lovi´c Determination of Capsaicin and Dihydrocapsaicin Content by HPLC Method in Products Purchased Online . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 478 E. Mlivo, A. Uzunovi´c, A. Baši´c-Halilovi´c, M. Daci´c, S. Pilipovi´c, and K. Duri´c
Contents
xiii
Artificial Intelligence and Machine Learning Explainable CNN-Based Cardiac Amyloidosis Classification from PET Images Through Manifold Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491 Lisa Anita De Santi, Filippo Bargagna, Maria Filomena Santarelli, Giuseppe Vergaro, Dario Genovesi, Michele Emdin, Assuero Giorgetti, and Vincenzo Positano Evaluation of Bacterial Biofilm Category Change Due to the Use of Different Signaling Molecules Using Random Forest Classifier . . . . . . . . . . . . 504 Abdullah Bjelak, Sara Deumi´c, Jasmin Kevri´c, and Monia Avdi´c Evaluating Bilateral Surface EMG Features for Automatic Identification of Gait Phase Transitions in Ground Walking Conditions . . . . . . . . . . . . . . . . . . . . 517 Francesco Di Nardo, Christian Morbidoni, Filippo Ventura, Alessandro Cucchiarelli, and Sandro Fioretti A Vital Signs-Driven Approach for Clustering the Responses to Diuretics Treatment in Premature Newborns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526 Riccardo Asnaghi, Nicolò Pini, Nimrod Goldshtrom, and Manuela Ferrario Identification of Atypical Cardiac Patterns Before and After Exercising Using Artificial Intelligence and Eulerian Video Magnification . . . . . . . . . . . . . . . 535 Carolina Ruiz-Laguado, Martha Lucia Zequera-Diaz, and Francisco Carlos Calderón-Bocanegra Clinical Natural Language Processing and Health Interoperability to Support Knowledge Management and Governance in Rare Cancers . . . . . . . . . 546 Laura Lopez-Perez, Eugenio Gaeta, Itziar Alonso, Alejo Esteban, Victor Gerardo Dominguez, Franco Mercalli, Elena Martinelli, Annalisa Trama, Maria Fernanda Cabrera, Maria Teresa Arredondo, and Giuseppe Fico Deep Learning Enabled Acute Ischemic Stroke Lesion Segmentation for Smart Healthcare Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553 Zhuldyz-Zhan Sagimbayev, Alisher Iglymov, Almagul Zhussupova, Meruyert Saifullakyzy, Doszhan Zhussupov, Dias Tashev, Gulden Zhanmukanbetova, and Raushan Myrzashova When an Explanation is not Enough: An Overview of Evaluation Metrics of Explainable AI Systems in the Healthcare Domain . . . . . . . . . . . . . . . . . . . . . . . 573 Essi Pietilä and Pedro A. Moreno-Sánchez
xiv
Contents
Liver Diseases Classification Using Machine Learning Algorithms . . . . . . . . . . . 585 Ivan Jovovi´c, Marko Grebovi´c, Lejla Gurbeta Pokvi´c, Tomo Popovi´c, ˇ c and Stevan Caki´ Interpretability and Personalization Aspects in the Development of Clinical Risk Assessment Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594 S. Paredes, S. Sousa, J. Henriques, T. Rocha, J. Sousa, and L. Gonçalves Ten Year Cardiovascular Risk Estimation: A Machine Learning Approach . . . . . 604 Dejan Babic, Luka Filipovic, Sandra Tinaj, Ivana Katnic, and Stevan Cakic Thyroid Hormones Parameter-Based Classification of Patient Health Status: An Analysis of Machine Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . 613 Zoja Š´ceki´c, Luka Filipovi´c, Ivana Katni´c, Nela Miloševi´c, and Stevan Šandi Artificial Intelligence-Based Ultrasound Imaging Classification for Infant Neurological Impairment Disorders: A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620 Lemana Spahi´c, Zerina Mašeti´c, Almir Badnjevi´c, Asim Kurjak, and Lejla Gurbeta Pokvi´c Cardiovascular Health in AI: A Comprehensive Overview to Acute Myocardial Infarction Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 628 Asja Muharemovi´c and Jasmin Kevri´c Sepsis Detection Using Features Extracted from Photoplethysmography . . . . . . . 636 Elena Adelucci, Martina Falagiani, Sara Lombardi, Piergiorgio Francia, and Leonardo Bocchi Digital Histopathological Discrimination of Label-Free Healthy Tissues by Decision Tree Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 647 José Luis Ganoza-Quintana, José Luis Arce-Diego, and Félix Fanjul-Vélez Heart Disease Prediction Using Logistic Regression Machine Learning Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654 Faris Hrvat, Lemana Spahi´c, and Amina Aleta Class Probability Distributions of a Neural Network Classifier of Multiple Sclerosis Lesions on Quantitative Susceptibility Mapping . . . . . . . . . . . . . . . . . . . 663 Šiši´c Nedim, Barakovi´c Muhamed, Almisreb Abd Ali, Granziera Cristina, and Rogelj Peter
Contents
xv
Forecasting Icterus with Machine Learning: An Advanced Classification Model Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673 Tamara Pavlovi´c, Marko Grebovi´c, Armin Alibaši´c, Milica Vukoti´c, and Stevan Šandi Evaluation of Deep Learning Techniques for Automatic Lesion Segmentation in Mammography Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684 Ivan Lazic, Niksa Jakovljevic, Milan Rapaic, Jasmina Boban, and Tatjana Loncar-Turukalo Integrating Machine Learning in Clinical Decision Support for Heart Failure Diagnosis: Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 696 Lemana Spahi´c, Adna Softi´c, Azra Durak-Nalbanti´c, Edin Begi´c, Bojan Staneti´c, and Haris Vrani´c The Security Framework: Determined and AI Impact? . . . . . . . . . . . . . . . . . . . . . . 706 Ivan Jovetic and Ivana Katnic The Future of (Individual) Security—Vision 2.0? . . . . . . . . . . . . . . . . . . . . . . . . . . 716 Ivan Jovetic, Milica Vukotic, and Sandra Tinaj Forecasting Meningitis with Machine Learning: An Advanced Classification Model Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725 Benjamin Dobardži´c, Armin Alibaši´c, Nela Miloševi´c, Bojana Mališi´c, and Milica Vukoti´c Analysis of Predictive Parameters in Prediction of the Occurrence of Type 2 Diabetes Using Machine Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . 732 Sumeja Hadžali´c, Arnela Obralija, Šeila Be´cirovi´c, Belma Pehlivanovi´c Kelle, and Ehlimana Krupalija Health Informatics, e-Health and Telemedicine Challenges and Outlook to Designing Cutting-Edge Mixed Reality Technologies of Human Pose Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743 Michela Franzò, Simona Pascucci, Franco Marinozzi, and Fabiano Bini Potentially Inappropriate Medications in Geriatric Patients with Type 2 Diabetes Mellitus: Practical Software Solution for Healthcare Professionals . . . . 755 Naida Omerovi´c, Anela Hadžifejzovi´c Trnka, Nermina Žiga Smaji´c, and Selma Škrbo Addressing Rapidly Aging Society Challenges Through Health 4.0 . . . . . . . . . . . 764 Pedro A. Moreno-Sánchez, Konstantinos Banitsas, Mark van Gils, and Maysam Abbod
xvi
Contents
A Virtual Reality-Based Setting to Investigate How Environments and Emotionally-Laden Stimuli Interact and Compete for Accessing Consciousness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773 A. Iannizzotto, S. Frumento, D. Menicucci, A. L. Callara, A. Gemignani, E. P. Scilingo, and A. Greco Smart Sensors for Daily-Life Data Collection Toward Precision and Personalized Medicine: The TOLIFE Project Approach . . . . . . . . . . . . . . . . . 783 Nicola Carbonaro, Marco Laurino, Alberto Greco, Carlotta Marinai, Francesca Giannetti, Francesca Righetti, Francesco Di Rienzo, Gianluca Rho, Lucia Arcarisi, Michele Zanoletti, Pasquale Bufano, Mario Tesconi, Nicola Sgambelluri, Danilo Menicucci, Carlo Vallati, and Alessandro Tognetti Exploring the Progress of Digital Transformation in the Healthcare System of Bosnia and Herzegovina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 795 Nerma Džaferovi´c, Adna Softi´c, Armin Elezovi´c, Faruk Guti´c, Haris Vrani´c, and Lejla Gurbeta-Pokvi´c Medical Engineering, Health Informatics and Medical Imaging (ME&HI) Profession and Curriculum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803 Nikitas N. Karanikolas, Mirjana Ivanovic, Alda Kika, Ridvan Alimehmeti, and Christos Skourlas Patients’ Health Information Exchange Within the European Union: A Croatian Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815 Hrvoje Belani, Tanja Bedovec, and Andreja Matkun Interaction Between Pharmaceutical Companies and the Public During the COVID-19 Pandemic—A Twitter Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 826 Sotirios Gyftopoulos, George Drosatos, Leandro Pecchia, Giuseppe Fico, and Eleni Kaldoudi Innovations in the Food System—Towards Next Generation of Multidisciplinary Initiatives The Many Faces of E. Faecium: From Probiotics to Pathogenesis . . . . . . . . . . . . . 837 Beatriz Daza-Prieto, Adriana Cabal-Rosel, Nadja Raicevic, Anna Stoeger, Johann Ladstaetter, Robert L. Mach, Werner Ruppitsch, and Aleksandra Martinovic
Contents
xvii
Antimicrobial Resistance in the Food Chain—Are We at the Point Where There is no Time to Wait? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845 Aleksandra Martinovic, Andrea Milacic, Nadja Raicevic, Amil Orahovac, Beatriz Daza, Marija Vugdelic, Adriana Cabal, and Werner Ruppitsch HPLC Q-TOF LC/MS Analysis of Inulin in Foods: Development of an Innovative Chromatography Method for Nutritional Enhancement . . . . . . . 856 Dado Latinovi´c Whole Genome Sequencing for Food Safety, Clinical and Public Health Microbiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 865 Adriana Cabal, Beatriz Prieto, Nadja Raicevic, Ariane Pietzka, Ali Chakeri, Patrick Hyden, Michael Kundi, Alexander Indra, Robert Mach, Julio Enrique Parra Flores, Aleksandra Martinovic, and Werner Ruppitsch Green Chemistry Analysis of Atmospheric Concentrations of PM 10 and PM 2.5 and Their Impact on the Health Status of the Population in Urban Industrial Cities of Bosnia and Herzegovina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 877 Aleksandra Babi´c, Nejla Piri´c, and Elma Imamovi´c Optimization of Glyphosate Adsorption Conditions on Pyrophyllite . . . . . . . . . . . 883 Tolic Tina, Klepo Lejla, Topcagic Anela, Copra-Janicijevic Amira, Omar Chahin, Kresic Dragan, and Ostojic Jelena Understanding Main Drivers of Global Decarbonization . . . . . . . . . . . . . . . . . . . . . 901 Ivana Vojinovic Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 911
Biomedical Signal Processing
Multimodal Registration Algorithm of Plantar Pressure Images and Digital Scanner for the Characterization of the Diabetic Foot in the BASPI/FootLab Laboratory Yeison Luna(B) and Martha Zequera Pontificia Universidad Javeriana, Bogotá, D.C, Colombia {yeison.luna,mzequera}@javeriana.edu.co
Abstract. This project focuses on the BASPI/FootLab research line, “Emerging Technologies To Support Health Care and Independent living”. This paper shows the continuity to the improvement of the DIAPETICS platform, which supports the prevention of ulcers in patients with diabetic foot developed by the BASPI/FootLab laboratory, the implementation of a multimodal image registration algorithm was carried out to analyze the distribution of plantar pressures with the morphology of the foot, using plantar pressure and digital scanner images of the laboratory, thus using image processing techniques that allow to visually characterize and understand diabetic foot patterns that provide new information towards early diagnosis and identification of patients with peripheral neuropathy considered in the model to be developed by the laboratory. Once the multimodal image registration algorithm was implemented, its performance was evaluated by comparing it with other methods that have been implemented in the literature, thus obtaining performance measurements of all the multimodal image processing techniques, that seem useful in the comparison and characterization of the diabetic foot to verify whether it is possible to determine a normal and abnormal plantar pressure threshold, according to the results obtained. Keywords: Elderly · Multimodal image registration · Plantar pressure · Podoscopy · Digital scanner · Diabetes · Peripheral diabetic neuropathy
1 Introduction Diabetes is one of the chronic diseases that currently affects 463 million people [1]. Within this population, a significant group of people presents one of its complications, called Peripheral Diabetic Neuropathy, causing the abnormal distribution of plantar pressures, which associated with poor vascularization leads to the presence of ulcers that can trigger future amputations of lower limbs [2]. The lack of education, information, and understanding of the manifestations of diabetic foot symptoms on the part of the patient and his family, contribute to the fact that the rate of amputees in the world grows even though there is a suitable technology for its diagnosis, which is not integrated into the clinical pathway of diabetes [2]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 3–12, 2024. https://doi.org/10.1007/978-3-031-49062-0_1
4
Y. Luna and M. Zequera
For this reason, to mitigate these problems, the purpose of this work is to find new technological tools for the diagnosis and support of pathway of the patient in diabetes, specifically in diabetic foot.
2 Methodology The project is divided into 4 stages: the first stage is the acquisition of plantar pressure images and digital scanner, a database was elaborated by using a protocol for the acquisition of these images in 12 volunteers of the BASPI/FootLab laboratory with the previous approval of the Research Committee and the Ethics Committee of the Faculty of Engineering. In the next stage, the pre-processing is done, which seeks to estimate a minimum rectangle that encloses the foot in each image to define three variables: scale, translation and rotation that will be the variables of the registration of the images. In the next stage, four techniques of multimodal image registration algorithms are implemented by varying the calculation of similarity: Mutual Information [5], Normalized Mutual Information, Cross Correlation and Normalized Cross Correlation, and for the final stage, the comparative evaluation of the techniques is made by calculating indicators such as the algorithm execution time, the Average Mutual Information of the Registration, the Average Sum of the squared intensity differences and the mean squared error. In the Fig. 1 it is shown the general block diagram of this work.
Fig. 1. General block diagram of the project
2.1 Protocol A dataset of 120 plantar pressure images and 50 scanner/electrical podoscope images was generated. For the acquisition of plantar pressure images, the Ecowalk pressure platform, from the company Ecosanit was used. The acquisition of images of the plantar morphology of the foot was performed with a biomedical device for clinical diagnosis called “Electronic Podoscope”. The podoscope used for the acquisition of the data is from the company INTERFI ´ SICA, and has the following characteristics: Before the acquisition of the data, the volunteers were clearly explained the conditions under which the acquisition will be performed, that it does not imply a risk and that they are free to withdraw at any time without any negative consequences for the volunteers. They were freely asked to sign an informed consent form, following the protocols of bioethics in human research recommended by the Research and Ethics Committee of the Faculty of Engineering of the Pontificia Universidad Javeriana.
Multimodal Registration Algorithm of Plantar Pressure Images
5
Prior to the acquisition of the data with the volunteers participating in the research, the personalized data required to create the different study cases were registered in the platform software, taking into account: the date, time and number of replicates that will be made in the acquisition of samples for the creation of the database, which will be anonymized under a reference code to avoid the identification of each volunteer, complying with the bioethical implications. The database of the volunteers participating in this study will be hosted in an institutional server where there is a space assigned for BASPI/FootLab laboratory research with restricted access and for research purposes. 2.2 Pre-Processing Pre-processing of the plantar pressure image In this section of the image preprocessing stage, the input is the plantar pressure image and the output is two pressure images separated by left foot and right foot, with a size of 455 x 455 pixels. Likewise, this sub-block has two internal blocks: Separation of left foot and right foot, and on the other hand, Edge replication. These internal blocks will be explained below. – Separation of left foot and right foot: The input to this internal block is the plantar pressure image, which is 455 x 455 pixels in size, in PNG format. The outputs of this internal block are two plantar pressure map images separated by foot. The internal process consists of cropping the input image, using software, into two images of size 228 x 455 pixels, one corresponding to the first 228 pixels of the columns of the original image, and another image corresponding to the last 228 pixels of the columns of the original image. – Edge replication: This section has as input the plantar pressure images of the left and right foot. The outputs of this block are the plantar pressure images of the left and right foot with a size of 455 x 455 pixels each. The internal process consists of replicating the outer edges, or adding a number of columns on the left and right side of the plantar pressure images of each foot, so the following operations are performed to obtain the number of columns to be added on each left and right edge: The input image has 228 columns and the output image has 455 columns. The difference of columns is calculated: 455 − 228 = 227 Next, the integer division of the columns is performed: 227 = 113.5 2 Therefore, 113 columns should be added to the left side and 114 columns to the right side of the plantar pressure images, thus obtaining a plantar pressure image for each foot and with a size of 455 x 455 pixels. Pre-processing of digital scanner/podoscope images In this section of the PreProcessing stage has as inputs the digital scanner images of the left and right foot,
6
Y. Luna and M. Zequera
and has as outputs the podoscope images with a size of 455 x 455 that will be the standard size to perform the multimodal image registration that will be explained in the next section. In addition, this sub-block has two internal blocks that will be explained below. – Perspective transformation: This part has as inputs the digital scan images of the left and right foot, and has as outputs the digital scan images of the left and right foot with a perspective transformation to eliminate the tilt of the image acquisition that may occur. The process consists of transforming the input image in such a way that the tilt effects are eliminated at the time of image acquisition. To do this, it is necessary to take the physical measurements of the glass area where each of the feet rests, and to calculate the aspect ratio of this area, as seen below. The physical measurements of height and height of the rectangular area has a width of 29.5 cm and 27.2 cm. With these measurements the calculation of the aspect ratio is made, which is given by the Eq. 1 AR =
w h
(1)
where AR is the aspect ratio, w is the width and h is the height. In this case, the aspect ratio is: AR =
29.5cm ≈ 1, 085 27.2cm
With this aspect ratio it is possible to calculate the desired output pixels of the image. For this purpose, we want to obtain an image with a width of 228 pixels, the procedure to find the height of the output image is presented below. The values of the width of the output image and the aspect ratio: wout = 228px; RA = 1.085 By solving the Eq. 1, the height of the output image is obtained: hout ≈ 210px In this way, the width and height data of the output image are obtained, so the output image will have a size of 228 x 210 pixels. – Edge Replication: As it was done for the plantar pressure images, this same procedure applies for the podoscope or digital scanner image, in this case we have input images of 228 x 210 pixels. The output images should be 455 × 455 pixels, which was defined as the standard size of the images for registration. So, the rows to be added at the top and bottom of the image are given by the following calculation: 455 − 210 = 245 The division by 2 is performed to distribute half of the rows to be added at the top and bottom edge of the image: 245 = 122.5 2
Multimodal Registration Algorithm of Plantar Pressure Images
7
In this way, 123 rows will then be added at the top and 122 rows at the bottom of the input images, to obtain images of size 455 x 455 pixels. On the other hand, obtaining the columns to add on the left and right side borders are given by the following calculation: 455 − 228 = 227 The division by 2 is performed to distribute half of the columns to be added on the left and right border of the image: 227 = 113.5 2 In this way, 113 columns will be added on the left edge and 114 columns on the right edge of the input images, to obtain images of size 455 x 455 pixels. Stimation of scale, translation and rotation This is the last section of the Preprocessing stage, where the variables of scale, translation and rotation are estimated to apply a geometric transformation prior to the plantar pressure image, which according to the theory of Image Registration and for the purposes of this project, will be the test image. This in order that the multimodal image registration algorithm has a shorter execution time, making a Pre-Registration. The inputs of this block will be the plantar pressure images of each foot, and the scan images of each foot. On the other hand, the outputs of this block will be the two preprocessed plantar pressure images with the geometric transformation using the estimated variables. – Minimum Rectangle: This section aims to find the minimum rectangle that encloses the area of the sole of the foot in the plantar pressure and scan images, in order to perform an estimation of scale, translation and rotation that must be applied to geometrically transform the plantar pressure image. The inputs of this block are the plantar pressure and podoscope images of both feet, and the outputs will be the pressure images with the geometric transformation of the estimated variables, scaled, translated and rotated. The procedure consists of taking the input images and mapping them into a grayscale, with this, a binarization must be applied for each of the images. Due to the noise that the binarization process can produce, a morphological opening and closing is applied to eliminate external noise and enclose all the morphological structure of the foot, and in this way find the shape of the foot in the pressure image and the scan image. Once the shape of the foot is found, the contour of the sole of the foot is found and the minimum area rectangle enclosing it is drawn. The rectangle found has three main attributes, which are: the center (x, y), the width and height (w, h) and finally, the angle of rotation theta. These will be the variables to be used in the estimation of transformation variables of the pressure image. – Stimation of Scale, Translation and Rotation Variables: This part consists of the estimation of the geometric transformation variables of the pressure image prior to being sent to the multimodal image registration algorithm block.
8
Y. Luna and M. Zequera
Once the minimum rectangle has been found in the plantar pressure and digital scanner images, the attributes of each of the rectangles of the plantar pressure and scanner images are used. For the estimation of the scale factor to be applied to the plantar pressure image, the expression (2) is used. S=
hpressure hscanner
(2)
where S is the scale factor, hpressure is the height of the rectangle enclosing the foot in the plantar pressure image, and hscanner is the height of the rectangle enclosing the foot in the scan image. For the estimation of the horizontal translation factor to be applied to the plantar pressure image, the expression (3) is used. X = Xpressure − Xscanner
(3)
where X is the horizontal translation factor, x pressure is the first component of the center of the rectangle enclosing the foot in the plantar pressure image, and x scanner is the first component of the center of the rectangle enclosing the foot in the scan image. For the estimation of the vertical translation factor to be applied to the plantar pressure image, it is used the expression (4) Y = Ypressure − Yscanner
(4)
where Y is the vertical translation factor, ypressure is the second component of the center of the rectangle enclosing the foot in the plantar pressure image, and yscanner is the second component of the center of the rectangle enclosing the foot in the scanner image. For the estimation of the rotation factor to be applied to the plantar pressure image, the expression (5) for the right foot is used. ϑ = ϑpressure − ϑscanner
(5)
where θ is the rotation factor, ϑpressure is the angle of rotation of the rectangle enclosing the foot in the plantar pressure image, and ϑscanner is the angle of rotation of the rectangle enclosing the foot in the scanner image. For the calculation of the rotation factor for the left foot, the expression is 6 ϑ = ϑscanner − ϑpressure
(6)
2.3 Multimodal Image Registration of Plantar Pressure Images and Scanner Images At this stage of the project, four multimodal image registration techniques are implemented. The general block diagram of the multimodal image registration algorithm is given by the one shown in the Fig. 2
Multimodal Registration Algorithm of Plantar Pressure Images
9
Fig. 2. General block diagram of image registration
The entries of the Multimodal Image Registration block are the previously preprocessed images. The podoscope and plantar pressure images. For the purposes of this project, a Multimodal Image Registration algorithm is proposed to estimate a geometric transformation for the plantar pressure image with four variables: Scale S, Horizontal translation (or translation on the x-axis) X, Vertical translation (or translation on the y-axis) Y, and finally, Rotation R. The optimization technique will be the multiscale search, that is, for each of the variables will be performed three cycles of search scales or search windows, each search window will be finer, so that the algorithm will converge to a specific value for each of the variables of the geometric transformation. The multi-scale search consists of dividing the reference image, which in this case is the podoscope image, into a specific number of initial intervals, where the variables of scale, horizontal and vertical translation, and rotation will be estimated. In this initial search window, S1, X1, Y1 and R1 values are estimated. Once the geometric transformations that satisfy the highest value of similarity measure are found, we proceed to the second cycle of the algorithm, in which a finer search is made, because it is divided again in the same number of intervals, starting from the position before the optimal one in the primary cycle and ending in a position after the one found in the first cycle, here the geometric transformations S2, X2, Y2 and R2 are estimated. Finally, the process is repeated again in the last cycle, so that the search is even finer and the last geometric transformation parameters that satisfy the best similarity measure are found, generating a multimodal image registration.
3 Results From the acquired data set, for the processing of the algorithm evaluation, the third images were chosen from the plantar pressure and podoscope images, in order to take the images that are farther away from the initial and final stage of the tests, where capture errors may occur due to the stabilization of the volunteer. The input images with which the algorithm was tested were a total of eighty, distributed as follows.
10
Y. Luna and M. Zequera
By image registration technique, there are twenty images, corresponding to four images per patient (left foot podoscope image, right foot podoscope image, left foot plantar pressure image and right foot plantar pressure image), times five, since this is the total number of volunteers to whom both types of images were taken. Since there are four techniques to be evaluated, the number of images per technique (20) is multiplied by a factor of four, which gives a total of eighty images. The comparison of metrics obtained in the implementation of the different multimodal image registration techniques was performed, the metrics to be used to verify the performance of each of the registers will be Mutual Information, Sum of squared intensity differences and Mean Squared Quadratic Error. Which are image similarity metrics based on the state of the art. The summary of the results are shown in the Table 1. The mutual information 9 is given by the entropy of one image 7, and by the joint entropy of two images 8. If this measure increases, the quality and performance of the multimodal registration will also increase. H (A) = ap(a)logp(a) (7) H (A, B) =
p(a, b)logp(a, b)
(8)
I (A, B) = H (A) + H (B) − H (A, B)
(9)
a
b
In the case of the sum of squared intensity differences, it should be taken into account that the lower the value of this metric, the better the quality of the recording. |I1 (x, y) − I2 (T (x, y))|2 SSD = (10) The mean square error (MSE), will be given by the Eq. 11 MSE =
N 1 (I1 (x, y) − I2 (T (x, y)))2 N
(11)
i=0
The execution time of the different logging techniques will also be a metric to determine the performance of the algorithm. The Table 1 shows the different metrics in the algorithms. The Table 2 shows the execution time of each algorithm.
4 Discussion The multimodal image registration system of plantar pressure imaging and podoscopy or digital scanning was implemented and evaluated. A proprietary data sample size was achieved, which although not of a significant size, showed important information due to the variety of plantar pressure patterns. The estimation of the geometric transformation parameters of the plantar pressure image requires a Pre-Registration, and a proper pre-processing of the data to have a
Multimodal Registration Algorithm of Plantar Pressure Images
11
Table 1. Average image registration evaluation metrics Technique
Average mutual information
Average Sum of squared intensity differences
Mean squared error
Mutual information
0,125
1,588E + 16
115,734
Normalized mutual Information
0,121
1,451E + 16
115,804
Cross correlation
0,113
2,518E + 16
115,527
Normalized cross correlation
0,121
1,306E + 16
115,879
Table 2. Execution times for each image registration technique Technique
Execution time [s ]
Mutual information
4,634
Normalized mutual information
5,329
Cross correlation
4,391
Normalized cross correlation
4,0609
better performance in the different image registration techniques. The rectangle found in this pre-processing section is an important result to be able to estimate size, location and rotation of the foot in an image. The multimodal image registration algorithm was implemented with state-of-the-art based techniques in order to make a comparison of performances and different evaluation metrics. According to the average mutual information of the different registrations evaluated, the best performing technique is the algorithm based on Mutual Information, being 1.03 times higher than the other image registration techniques. This indicates that the quality of the registration has a higher performance. According to the average of the sum of squared intensity differences, indicating that the lower the value, the better the quality of the registration, the technique based on Normalized Mutual Information has the lowest value among the other techniques. According to the mean square error, the multimodal image registration technique with the best performance in this metric was the Cross Correlation based, while the Normalized Cross Correlation based registration technique had the highest mean square error of all the other registration techniques. In terms of execution times, the Normalized Cross Correlation-based multimodal image registration technique takes the least time to execute, while the Normalized Mutual Information-based technique has the longest execution time, with 5.33 s of execution.
12
Y. Luna and M. Zequera
In conclusion on the performance of the different multimodal image registration techniques, taking into account the different image registration quality metrics and execution times, the best technique is the one based on Mutual Information.
5 Conclusion A sample size of our own data was achieved, which although not of a significant size, showed important information due to the variety of plantar pressure patterns. Performing a pre-registration, using morphological operations on binarized images to find a minimum rectangle to enclose the foot, and performing the estimation of geometric transformation parameters S,X,Y,R allows the registration algorithm to be more efficient, since it starts from a first approach to the image registration. For future work it is suggested to replace the electronic podoscope with a specialized 2d scanner system for plantar surface morphology imaging, as the images taken from the podoscope will be affected by different environmental factors. It is recommended to increase the sample size to be able to perform a more exhaustive evaluation of the different multimodal image registration techniques implemented, and in the same way, to test the designed algorithm with images of another type, for example, heat maps of the foot. For the improvement of the execution times, it is suggested to review different code optimization techniques, in order to have results in less time.
Conflict of Interest. The authors declare that they have no conflict of interest.
References 1. International Diabetes Federation: Worldwide Toll of Diabetes (2019). https://diabetesatlas. org/en/sections/worldwide-toll-of-diabetes.html 2. International Diabetes Federation: IDF Diabetes Atlas (2019) 3. Zitová, B., Flusser, J.: Image registration methods: a survey. 21, 977–1000 (2003). https://doi. org/10.1016/S0262-8856(03)00137-9 4. Zulkifli, S.S., Loh, W.P.: A state-of-the-art review of foot pressure. 26, 25–32 (2020). https:// doi.org/10.1016/j.fas.2018.12.005 5. Woo, J., Stone, M., Prince, J.L.: Multimodal registration via mutual information incorporating geometric and spatial context. 24, 757–769 (2015) 6. Oliveira, F.P., Tavares, J. M.: Registration of plantar pressure images. Int. J. Numer. Methods Biomed. Eng. 28(6–7), 589–603 (2012) lR: 20160511; CI: Copyright © 2011; JID: 101530293; OTO: NOTNLM; 2011/03/28 00:00 [received]; 2011/06/13 00:00 [revised]; 2011/06/16 00:00 [accepted]; 2014/11/04 06:00 [entrez]; 2012/06/01 00:00 [pubmed]; 2015/07/07 06:00 [medline]; ppublish
Semi-Automated Calculation of Baroreflex Sensitivity (BRS) Indices Magdalena Krbot Skori´c1,2(B) , Ivan Adamec1,3 , Ivan Moštak3 , Nika Višnji´c3 , Mario Cifrek2 , and Mario Habek1,3 1 Department of Neurology, Referral Center for Autonomic Nervous System Disorders,
University Hospital Center Zagreb, Zagreb, Croatia [email protected] 2 Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia 3 School of Medicine, University of Zagreb, Zagreb, Croatia
Abstract. The aim of this research was to objectify the method of semiautomatized calculation of the baroreflex sensitivity (BRS) indices. In sixty-nine individuals with a history of vasovagal syncope and no other neurological or systemic illnesses (mean age 47.04 ± 11.18, 55 females) blood pressure (BP) response to Valsalva maneuver (VM) was performed and BRS indices were calculated from the systolic BP curves (sBP): BRSa1, alpha BRSa (α-BRSa) and beta BRSa (βBRSa). Semi-automatizated software was created according to previously known formulas, with additional calculation of average baseline BP values. Median values for manually calculated indices were 21.45 for BRSa1, 7.01 for α-BRSa, and 1.37 for β-BRSa. For semi-automatically calculated indices values were 23.91 for BRSa1, 6.99 for α-BRSa, and 1.19 for β-BRSa. There was a statistically significant correlation between the manually calculated and semiautomatically generated results for all three coefficients (BRSa1: rs = 0.920, p < 0.001; α-BRSa: rs = 0.879, p < 0.001; β-BRSa: rs = 0.889, p < 0.001). Automatization of BRS indices calculation shows results consistent and highly corelated with manually calculated indices, reduces time required for calculation and reduces the impact of subjective human component on the calculations. Method for semi-automated calculation of BRS indices enhances the speed and precision of the calculation rendering the process more pragmatic for future studies. Keywords: Baroreflex sensitivity · Autonomic nervous system · Valsalva maneuver
1 Introduction Whenever a person stands up from a supine position around 800 ml of blood tends to pool from the head to the splanchnic circulation and the deep veins of the legs due to the gravity [1]. If not counteracted, it may lead to the orthostatic hypotension and syncope. The main reflex that enables the maintenance of equal blood pressure in the supine and upright position is the baroreflex. It is mediated through mechanoreceptors in carotid arteries and the aortic branch that detect blood pressure changes (BP) and by activating © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 13–21, 2024. https://doi.org/10.1007/978-3-031-49062-0_2
14
M. Krbot Skori´c et al.
autonomic centers in the brainstem proceed to modulate adrenergic and vagal activity on peripheral resistance and cardiac output [1]. The most common method of testing baroreflex function is the measurements of the BP response to the Valsalva maneuver [2]. When Valsalva maneuver is performed, a fall in blood pressure (BP) occurs, resulting in baroreflex activation leading to a sympathetically driven increase in peripheral resistance which ultimately leads to BP recovery [3]. Through detailed analysis of the BP response, baroreflex sensitivity (BRS) indices can be calculated providing further information about the baroreflex adrenergic and vagal components [4]. Another challenge in assessing components of the Valsalva maneuver is the detection of the response type, which according to Palamarchuk et al. can be divided into a balanced, augmented and suppressed response [5]. Although BRS indices are an important part of the baroreflex evaluation, they are usually calculated manually, which is time consuming process and are dependent on the subjective assessment hence prone to human error [4]. The aim of this study was to objectify the method by semi-automatization of the calculation of BRS indices.
2 Methods 2.1 Participants This was a retrospective analysis of results of autonomic nervous system (ANS) testing in participants referred to the Laboratory for testing of the ANS, the Referral Center for the Autonomic Nervous System Disorders of the Department of Neurology, University Hospital Center Zagreb, Croatia. Sixty-nine participants with a history of vasovagal syncope and no other neurological or systematic illness were included. Patients receiving medications with known influence on the ANS were excluded from the analysis. All participants signed informed consent for the ANS testing and the study was approved by the Ethical Committee of the University Hospital Center Zagreb. 2.2 Autonomic Nervous System Testing ANS testing was performed in the Referral Center for the Autonomic Nervous System Disorders, with Task Force Monitor (TFM) (CNSystemsMedizintechnik AG, Austria). Signals were continuously measured beat-to-beat and exported in.mat format for the purpose of semi-automated analysis. The test protocol included deep breathing test, Valsalva maneuver (VM) and tilt-up table test [2]. For the purpose of this study only participants in whom blood pressure (BP) response to VM was reproducible, were included in the final analysis. VM was performed in the supine position, by the participants blowing into the tube connected to a mercury manometer, so that the pressure on the manometer is maintained at 40 mm Hg for 15 s. The tube had a small air leak to prevent closing of the glottis. The test was repeated until a reproducible response was obtained and the BP curves on visual inspection allowed measurements of the BP changes [6]. The following parameters were manually calculated from the mean BP curve: maximal drop of the mean BP during phase II (phase IIE) compared to the level before the start of the test, the peak of the mean
Semi-Automated Calculation of Baroreflex Sensitivity
15
BP at the end of late phase II (IIL—recovery) and overshoot in the phase IV. Pressure recovery time (PRT) was calculated from the systolic BP curve (sBP). BRS indices were calculated according to presented formulas in two ways: manually, by one of the authors experienced in ANS testing, and through specifically developed software for the analysis. In both cases, the identical formulas for BRS calculation were used, the only difference was in defining the baseline BP level. For the manual calculation, the baseline BP level was defined with visual inspection of the BP curve in the period before the start of the VM. For the semi-automated calculation, the baseline BP level was defined as average value of the last minute before the start of the VM. 2.3 Manual Calculation of the Baroreflex Sensitivity Analysis During the VM, three types of systolic BP curves were identified according to previously published data: BAR—balanced autonomic response; AAR—augmented autonomic response; SAR—suppressed autonomic response [5]. For each participant, the indices of BRS from the systolic BP curves during the VM were calculated (BRSa1, α-BRSa, and β-BRSa). BRS indices were calculated according to previously published formulas [4]. 2.4 Semi-Automated Calculation of the Baroreflex Sensitivity Analysis Using MATLAB R2019b program package, a application has been developed for evaluation of adrenergic and vagal BRS components in the VM. The process started with loading of the signal and accompanying parameters in.mat format followed by choosing the best Valsalva maneuver for the analysis, which is shown in the Fig. 1. Signals were checked for consistency and interpolated with cubic spline method for any missing data. Since only sBP was of interest for this analysis, HR analysis was not in focus, therefore premature ventricular contractions or arrhythmias were not further investigated. It is required to manually assign points in systolic blood pressure (sBP) to reconstruct the maneuver correctly and properly define phases (Fig. 2), so parameters of the sBP can be calculated. Entirely automated process was not considered due to possible existing dysautonomia and multiple variations in sBP response to VM, which makes automated calculation prone to errors (including protocol errors if the patient was not co-operative or could not perform the task).
16
M. Krbot Skori´c et al.
Fig. 1. Process of choosing best Valsalva maneuver. Table on the left – available parts of ANS recording, along with their duration. Table on the right - Valsalva maneuvers available for analysis.
Fig. 2. Reconstruction of Valsalva maneuver. Maneuver is reconstructed after entering correct local minimums and maximums of the sBP curve (if possible). If any of the four phases are missing, another Valsalva maneuver must be chosen.
Semi-Automated Calculation of Baroreflex Sensitivity
17
With semi-automated calculation, extreme pathology and incomplete straining test can be excluded from further analysis by user. Consistency of straining purely depends on the operator navigating the patient during test. sBP parameters of interest, necessary for further analysis, are shown in Fig. 3.
Fig. 3. Systolic blood pressure parameters. Baseline – mean value of sBP 60 s before start of Valsalva maneuver; valley – lowest sBP value in phase II; rise – increase of sBP in late phase II; overshoot – sBP peak in phase IV; A- maximal drop of the sBP during phase IIE compared to the level before the start of the test; B - maximal drop of the sBP during phase III compared to the peak of the sBP at the end of phase IIL; C – an increment from the end of phase IIE to the end of phase IIL; D – sBP recovery during PRT; E – an increment of sBP after reaching baseline in phase IV; t_α – late phase II duration; PRT – pressure recovery time, time required for sBP to reach baseline in phase IV.
18
M. Krbot Skori´c et al.
Features relevant to the research, such as: BRSa1, α-BRSa, β-BRSa and type of the curve pattern in response to VM are extracted from the data and presented in the numerical format. Adrenergic index BRSa1 was defined as sum of the amplitude of BP fall in early phase II (sBP parameter A) and three-fourths of the amplitude of BP fall in phase III (sBP parameter B) divided by pressure recovery time (PRT), [4, 5] according to: BRSa1 =
A + 0.75 ∗ B PRT
Discrete adrenergic indices reflect sBP recovery from drop in late phase II as α-BRSa (sBP parameter C—an increment of sBP from the end of phase IIE to the end of phase IIL sBP, t_α – late phase II duration) and adrenergic response from sBP drop in phase IV as β-BRSa (sBP parameter D—sBP recovery during PRT) [4]. If sBP did not fall below baseline in phase III, PRT and D do not exist, therefore BRSa1 and β-BRSa cannot be evaluated. tα ∗ C α − BRSa = tα2 + C 2 PRT ∗ D β − BRSa = √ PRT 2 + D2
2.5 Statistical Analysis The Kolmogorov-Smirnov test was applied to see whether the data have a normal distribution. Differences in quantitative variables were determined with the use of the parametric paired t-test and non-parametric Wilcoxon test. Correlations were tested with Pearson’s and Spearman’s correlation methods. P values less than 0.05 were considered as significant. Software used for statistical analysis was IBM SPSS, version 25.
3 Results Sixty-nine participants (mean age 47.04 ± 11.18, 55 females) were included in analysis. For manual calculation, BRSa1 and β-BRSa were available for 53 participants; and for semi-automated calculation, BRSa1 and β-BRSa were available for 49 participants. αBRSa was available for all participants in both forms of calculation. Descriptive values for BRS indices are presented in Table 1. There was statistically significant correlation between the manually calculated and semi-automatically generated results for all three coefficients (BRSa1: rs = 0.920, p < 0.001; α-BRSa: rs = 0.879, p < 0.001; β-BRSa: rs = 0.889, p < 0.001). There was no statistically significant difference in BRSa1 and α-BRSa indices between two forms of calculation (p-values > 0.05), while there was statistically significant difference between two forms of calculations for the β-BRSa index (p = 0.005).
Semi-Automated Calculation of Baroreflex Sensitivity
19
Table 1. Descriptive data for manually and semi-automated calculated BRS indices
Semi-automated
Manual
Mean
Median
St. Dev
Min
Max
BRSa1
33,53
23,91
24,03
4,62
95,74
α-BRSa
6,90
6,99
2,27
0,08
11,57
β-BRSa
1,58
1,19
1,22
0,06
4,70
BRSa1
35,78
21,45
37,09
4,97
206,38
α-BRSa
6,88
7,01
2,49
0,67
14,69
β-BRSa
1,78
1,37
1,31
0,20
5,14
For manual calculations, 21 participants had BAR type of curve, 8 had SAR type of curve and 39 had AAR type of curve. For semi-automated calculations, 18 participants had BAR type of curve, 9 had SAR and 40 had AAR type of curve. As presented in Table 2, 8 participants were characterized with different types of curves. Table 2. Comparison of different types of curves for manual and semi-automated calculations Manual Semi-automated
BAR
SAR
AAR
BAR
18
0
0
SAR
0
5
4
AAR
1
3
35
BAR - balanced autonomic response; AAR - augmented autonomic response; SAR - suppressed autonomic response
4 Discussion This study provides a method for semi-automatized calculation of BRS indices with statistically significant strong correlation between the manually calculated and semiautomatically generated results for BRSa1, α-BRSa and β-BRSa. Although there was no statistically significant difference in BRSa1 and α-BRSa indices, there was statistically significant difference between two forms of calculations for the β-BRSa. This may be due to the different means of calculating BP baseline values and secondarily affecting PRT. Namely, the beta-adrenergic component is dependent on the pressure recovery of BP values from the fall in phase IV to reaching baseline values. While baseline BP level was defined with visual inspection of the BP curve in the period before the start of the VM for manual calculation, the semi-automated process defined the baseline BP as the average value of the last minute before the start of the VM. This disparity might have led to the difference of β-BRSa values between the manual and semi-automatized
20
M. Krbot Skori´c et al.
values. This is of particular importance in studies comparing BRS indices in two different timepoints, where manual calculation may give differences which are consequence of human error. Similarly, eight participants were found to have different types of VM curves as defined by Palamarchuk et al. [5]. The decision for the type of the curve was again mostly dependent on baseline BP value, and according to visual inspection, for these participants the BP baseline curve was covered with noise. Superimposed noise makes it difficult to detect the real baseline value in the manual type of calculation, and it contributes to the difference in determination of the curve types. Therefore, the differences between manual and semi-automatized calculations for BRS indices and type of VM curves might reflect the superiority of semi-automated objective determination of baseline BP values, as opposed to the subjectiveness of visual inspection. The values of rapid and precise BRS indices calculation are manifold. BRS indices correlate well with signs of autonomic failure such as orthostatic hypotension [3]. Furthermore, changes in adrenergic BRS increase with greater severity and distribution of autonomic failure thus providing a possible marker for the progression of the disease [7]. It has been demonstrated that patients with heart disease frequently have abnormal BRS and its presence represents a risk of cardiovascular mortality [8]. As well, patients with neuroimmunological diseases, such as multiple sclerosis, experience adrenergic hyperactivity which predisposes them to vascular comorbidities [9]. Therefore, calculation of BRS sensitivity can be useful in a clinical setting with a prognostic value of future cardiovascular risk. Moreover, BRS calculation may prove to be useful in the setting of therapeutic development. Studies have shown that the baroreflex could represent a target of treatment to reduce the negative impact of sympathetic hyperactivity by the means of electric stimulation [10]. Such research would require a great number of patients which would necessitate semi-automated BRS calculation as provided in the current study. In conclusion, this study provides a method for semi-automated calculation of BRS indices that enhances the speed and precision of the calculation rendering the process more pragmatic for future studies.
References 1. Freeman, R.: Clinical practice. Neurogenic orthostatic hypotension. N. Engl. J. Med. 358(6), 615–624 (2008) 2. Freeman, R.: Assessment of cardiovascular autonomic function. Clin. Neurophysiol.: Off. J. Int. Fed. Clin. Neurophysiol. 117(4), 716–730 (2006) 3. Novak, P.: Assessment of sympathetic index from the Valsalva maneuver. Neurology 76(23), 2010–2016 (2011) 4. Palamarchuk, I.S., Baker, J., Kimpinski, K.: Non-invasive measurement of baroreflex during Valsalva maneuver: evaluation of alpha and beta-adrenergic components. Clin. Neurophysiol.: Off. J. Int. Fed. Clin. Neurophysiol. 127(2), 1645–1651 (2016) 5. Palamarchuk, I., Baker, J., Kimpinski, K.: Non-invasive measurement of adrenergic baroreflex during Valsalva maneuver reveals three distinct patterns in healthy subjects. Clin. Neurophysiol.: Off. J. Int. Fed. Clin. Neurophysiol. 127(1), 858–863 (2016) 6. Novak, P.: Quantitative autonomic testing. J. Vis. Exp.: JoVE (53), 2502 (2011) 7. Schrezenmaier, C., Singer, W., Swift, N.M., Sletten, D., Tanabe, J., Low, P.A.: Adrenergic and vagal baroreflex sensitivity in autonomic failure. Arch. Neurol. 64(3), 381–386 (2007)
Semi-Automated Calculation of Baroreflex Sensitivity
21
8. De Ferrari, G.M., Sanzo, A., Bertoletti, A., Specchia, G., Vanoli, E., Schwartz, P.J.: Baroreflex sensitivity predicts long-term cardiovascular mortality after myocardial infarction even in patients with preserved left ventricular function. J. Am. Coll. Cardiol. 50(24), 2285–2290 (2007) 9. Habek, M., Mutak, T., Nevajdi´c, B., Puci´c, D., Crnošija, L., Krbot Skori´c, M.: Adrenergic hyperactivity: a missing link between multiple sclerosis and cardiovascular comorbidities? Acta Neurol. Belg. 120(3), 581–587 (2020) 10. La Rovere, M.T., Pinna, G.D., Raczak, G.: Baroreflex sensitivity: measurement and clinical implications. Ann. Noninvasive Electrocardiol.: Off. J. Int. Soc. Holter Noninvasive Electrocardiol. Inc. 13(2), 191–207 (2008)
Preliminary Assessment of the Samsung Galaxy Watch 5 Accuracy for the Monitoring of Heart Rate and Heart Rate Variability Parameters Gianluca Rho1,2(B) , Francesco Di Rienzo1 , Carlotta Marinai1 , Francesca Giannetti1 , Lucia Arcarisi1 , Pasquale Bufano3 , Michele Zanoletti3 , Francesca Righetti1 , Carlo Vallati1 , Marco Laurino3 , Nicola Carbonaro1,2 , Alessandro Tognetti1,2 , and Alberto Greco1,2 1 Dipartimento Di Ingegneria Dell’Informazione, University of Pisa, Largo Lucio Lazzarino 1,
56122 Pisa, Italy [email protected] 2 Research Center ”E. Piaggio”, University of Pisa, Largo Lucio Lazzarino 1, 56122 Pisa, Italy 3 Institute of Clinical Physiology, National Research Council, Via Giuseppe Moruzzi 1, 56124 Pisa, Italy
Abstract. In the last years, commercial smartwatches have gained popularity as non-invasive and wearable devices to be exploited for the monitoring of the cardiovascular system in daily-life settings. However, their reliability is still unclear. In this preliminary study, we evaluated the accuracy of heart rate (HR) and HR variability (HRV) estimates obtained from the Samsung Galaxy Watch 5 (SGW5) compared to a common research-grade ECG sensor, i.e., the Shimmer3 ECG unit (ShimECG), during both a resting and walking conditions. For each condition, we compared HRV features of SGW5 and ShimECG extracted in time, frequency, and non-linear domains through correlation and Bland-Altman analysis. Additionally, we compared SGW5 performance with those obtained from a research-grade PPG sensor. Our results revealed an unbiased and high-quality estimate of mean HR obtained from the SGW5. Moreover, at rest, other relevant HRV features showed a significant correlation between the SGW5 and ShimECG. Conversely, during the walking condition, we found poor performances for both PPG devices for most of the HRV features. Such preliminary results confirm the reliability of SGW5 to estimate mean HR. However, the reliability of SGW5- derived PRV to extract sympathovagal correlates is still an open question and deserves further investigation. Keywords: Wearable devices · Heart rate monitoring · Heart rate variability · Smart-watch
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 22–30, 2024. https://doi.org/10.1007/978-3-031-49062-0_3
Preliminary Assessment of the Samsung Galaxy Watch …
23
1 Introduction Heart rate (HR) and heart rate variability (HRV) are physiological parameters reflecting the general well-being of a subject [20]. Variations of such parameters can be adopted to evaluate stressful conditions, anxiety and panic [17]. Furthermore, they allow to identify autonomic imbalances associated with cardiovascular and respiratory dysregulation [14], as well as psychiatric disorders [3]. Electrocardiography (ECG) is the gold standard method adopted to estimate HR and HRV parameters in clinical and research settings. However, it is hardly used in dailylife contexts where free movements are required. To address this issue, wearable ECG sensors including shirts [13] and chest belts [9, 15] have been proposed. However, they can be obtrusive and uncomfortable over day-long recording periods. An alternative to ECG to estimate HR/HRV is the use of photoplethysmography (PPG) sensors. The pulse rate variability (PRV) estimated from the PPG signals can be used as a reliable surrogate of the HRV [12, 16]. Specifically, PPG estimates the volume variation of superficial blood vessels at either the fingertip or the ear lobe through an LED and a photo-receiver [8]. Although clinical-grade PPG monitoring devices significantly limit movements due to the constraint of wires, recently, wearable PPG sensors have rapidly spread as a key equipment of most commercial smartwatches (SW). Such large diffusion, combined with their ease of use and connectivity features with other devices, as well as the sensor positioning on the wrist, have led to the widespread use of SWs also in scientific research applications [2, 6, 9, 10, 19]. In the last years, several SW applications have been tested by comparing the SW performance against medical-grade ECG and PPG, e.g., the Apple Watch [11], Empatica E4 [13], Fitbit Charge HR [21], and Microsoft band 2 [11]. Particularly, they showed accurate HR estimates during resting conditions, with decreasing performances at the increase of the subjects’ activity intensity. However, on the one hand, such devices do not always provide direct access to raw data, and HR information is extracted through blackbox proprietary algorithms. On the other hand, some of them have a high cost that limits their spread. These limitations are potentially overcome by the open-source Android Wear operating system (OS), which has been recently adopted by several commercial SWs. Particularly, Android Wear OS allows developers and researchers to access sensors’ data through the development of custom applications. In this study, we evaluate the performances of an Android Wear OS SW, i.e., the Samsung Galaxy Watch 5 (SGW5) [1], on a group of healthy subjects during a resting state and a walking condition. The SGW5 relative low-cost and its long-life battery [1] could make it suitable for carrying out 24h HR and HRV monitoring research studies on a consistent number of subjects. The evaluation is performed against the measurements obtained from a validated ECG device, i.e., the Shimmer3 ECG unit [7]. Furthermore, we compare the SGW5 performances with those obtained by a widely used researchgrade PPG device, i.e., the Shimmer3 GSR + unit. For each PPG signal, we derive HRV parameters in the time, frequency, and non-linear domains, respectively, and we evaluate
24
G. Rho et al.
Pearson correlation coefficients with the same estimates obtained from ECG. Finally, we evaluate the device accuracy through a Bland-Altman analysis.
2 Materials and Methods 2.1 Subjects The study was conducted according to the guidelines of the Declaration of Helsinki. Twenty healthy volunteers (age 39 15, 7 females) signed an informed consent to take part in the study. Subjects self-reported no history of cardiovascular diseases. 2.2 Experimental Settings The experimental protocol consisted of two distinct conditions: (1) 1 min of resting state seated on a chair, with the arms resting on the table, and (2) 1 min of walking. Subjects were asked not to talk during the experiment. Moreover, during (1), subjects were asked to minimize movements. The order of the conditions was randomized across subjects. During each condition, we acquired the PPG signal at the wrist of the subject’s nondominant hand through the SGW5. We developed an ad-hoc Wear OS application to retrieve raw data from the PPG sensor, together with the relative universal timestamp, at the maximum available sampling frequency of 25 Hz. The data was sent via Bluetooth communication to a smartphone and then stored in a computer for the processing stage. We acquired the ECG signal through the Shimmer3 ECG unit (SHIMMER research, Dublin, Ireland) as the gold standard for this study [7]. The sensor was mounted on the subjects’ chest through an elastic band, with four leads placed on the left arm, right arm, left leg, and right leg, according to the manufacturer’s guidelines. We recorded ECG as the difference between the right leg and left arm leads, at the sampling frequency of 400 Hz. As an additional term of comparison, we acquired the PPG signal by using also the Shimmer3 GSR + unit (shimPPG). The signal was acquired at the tip of the first finger of the subjects’ non-dominant hand, at the sampling frequency of 25 Hz. To facilitate the following data processing, we synchronized the recordings of the SWG5, ECG, and shimPPG. 2.3 Data Processing The raw data were preprocessed in Matlab (Version R2021b, Mathworks, USA). We filtered both the SGW5 and shimPPG data in the (0.7–1.8)Hz frequency range through a zero-phase band-pass IIR filter (transition band = 0.1 Hz). We identified pulse peaks in the signals through the multi-scale peak and trough detection (MSPDT) algorithm [4]. The outcome of the MSPDT algorithm was visually inspected, and missing
Preliminary Assessment of the Samsung Galaxy Watch …
25
peaks or peaks not properly identified were manually corrected where possible. Afterwards, we imported peak-to-peak (P-P) distance time series in Kubios HRV [20], and we derived PRV signals after uniform interpolation at 4 Hz. The PRV was further corrected for the presence of artifacts (e.g., ectopic beats, abnormal P-P values) through the automatic artifact correction algorithm using a conservative threshold (low option in Kubios; see [20] for details). The procedure was applied to both SWG5 and shimPPG data. Regarding ECG, we applied a zero-phase band-pass IIR filter (transition band = 0.1 Hz) in the (0.7–25)Hz range. We estimated HRV time series using Kubios through the automatic QRS complex detection algorithm [20], followed by a uniform interpolation at the sampling frequency of 4 Hz. HRV artifacts were corrected through the Kubios automatic artifact correction algorithm using the same threshold as for the PRV signals. Using Kubios, we extracted the following features from PRV and HRV time series: (1) meanRR (i.e., the mean distance between consecutive peaks); (2) stdRR (i.e., the standard deviation of RR intervals); (3) RMSSD (i.e., root mean squared differences of successive RR intervals); (4) LF (i.e., low-frequency band; (0.04–0.15)Hz); (5) HF (i.e., high-frequency band; (0.15–0.40)Hz); (6) LF/HF ratio; (7) SD1, SD2 (i.e., the standard deviations of Poincaré plot). 2.4 Statistical Analysis For each of the two experimental conditions (i.e., resting-state, walking) and for each of the features extracted (see Sect. 2.3), we computed the Pearson correlation coefficient between the SGW5 and the ECG, and between the shimPPG and ECG, to evaluate the overall degree of agreement among measurements. The resulting p-values were corrected for multiple comparisons between different devices (i.e., SGW5 vs ECG, shimPPG vs ECG) with the Bonferroni method. Moreover, we conducted an individual Bland-Altman (BA) analysis [5] on each SGW5 feature to evaluate its bias and precision with respect to the ECG estimates. The same statistical procedure was applied to the shimPPG features. The presence of a significant bias was assessed through a paired t-test on the difference between devices’ measurements against the null-hypothesis of no difference. P-values were corrected for multiple comparisons. 2.5 Results In Table 1, we report the Pearson correlation coefficient ρ between SGW5 and ECG, and between shimPPG and ECG, for each of the estimated features (see Sect. 2.3), during both the resting state and walking condition.
26
G. Rho et al.
Table 1. Pearson correlation coefficient between the PRV and HRV features during the resting state and walking condition. Significant correlations are highlighted in bold and marked by * (*:p < 0.05, **:p < 0.01, ***p < 0.001; p-values adjusted with the Bonferroni correction). SGW5 versus ECG shimPPG vs ECG Feature
Rest
Walk
Rest
Walk
MeanRR
0.99***
0.63**
0.99***
0.69***
stdRR
0.69***
0.28
0.61**
0.41
RMSSD
0.49
0.37
0.34
0.34
LF
0.74***
0.05
0.96***
0.12
HF
0.64**
0.09
0.55*
0.45
LF/HF
0.47
0.14
0.61**
0.36
SD1
0.49
0.37
0.34
0.34
SD2
0.85***
0.22
0.87***
0.48
We observed similar correlations for the SGW5 and shimPPG at rest. Specifically, meanRR estimates showed an almost perfect degree of agreement with ECG ones (ρ SGW 5vsECG = 0.99, ρ shimPPGvsECG = 0.99). Moreover, SGW5 and shimPPG performances were comparable for both stdRR (ρ SGW 5vsECG = 0.69, ρ shimPPGvsECG = 0.61) and SD2 (ρ SGW 5vsECG = 0.85, ρ shimPPGvsECG = 0.87). In the spectral domain, we observed significant correlations for both LF and HF power estimates. Particularly, SGW5’s LF measurements showed a high correlation with those obtained from ECG (ρ SGW 5vsECG = 0.74), although shimPPG correlation was even higher (ρ shimPPGvsECG = 0.96). Conversely, SGW5’s HF estimates showed a higher correlation with ECG, compared to shimPPG (ρ SGW 5vsECG = 0.64, ρ shimPPGvsECG = 0.55). However, while we found a significant correlation between shimPPG and ECG for the LF/HF power ratio (ρ shimPPGvsECG = 0.61), such a relationship was not observed for the SGW5. Finally, we did not find a significant correlation for both RMSSD and SD1 estimates obtained from SGW5 and shimPPG. During the walking condition, the overall agreement between PRV and HRV features worsened with respect to the resting state for both SGW5 and shimPPG. Devices showed comparable significant correlation coefficients for meanRR, with respect to ECG (ρ SGW 5vsECG = 0.63, ρ shimPPGvsECG = 0.69). Nevertheless, none of the other features showed a significant correlation. In Figs. 1, 2, we report the results of the BA analysis for both the SGW5 and shimPPG against ECG, during the resting state and walking condition, respectively. Particularly, we report only those features for which we observed a significant correlation between SGW5 and ECG. The SGW5 estimated meanRR during resting state with no bias, while shimPPG overestimated meanRR by 9 ms (Fig. 1). Conversely, the SGW5 showed a higher negative bias (i.e., an overestimation), compared to shimPPG, for the estimation of stdRR (SGW5 = −20 ms, shimPPG = −8 ms; Fig. 1b), HF (SGW5 = −0.001 ms2 , shimPPG = 0 ms2 ; Fig. 1d), and SD2 (SGW5 = −0.007, shimPPG = 0; Fig. 1e). Particularly, both devices showed a tendency to overestimate more HF
Preliminary Assessment of the Samsung Galaxy Watch …
27
power with the increase of its average magnitude (see Fig. 1d). On the contrary, devices showed a tendency to underestimate more the LF/HF ratio with the increase of its average magnitude, with the SGW5 having a greater bias, compared to the shimPPG (SGW5 = 2.402, shimPPG = 2.008). a)
Mean RR
d)
b)
HF
Std RR
e)
c)
LF
SD2
Fig. 1. BA analysis results for the SGW5vsECG and shimPPGvsECG comparisons of a) meanRR (ms), b) stdRR (ms), c) LF (ms2 ), d) HF (ms2 ), e) SD2 (ms) PRV/HRV features during restingstate. For each comparison and for each feature, we report the bias, computed as the average difference between the devices’ measurements, and the limits of agreement, indicating the 1.96 standard deviation interval around the bias. Biases that significantly differed from 0 are indicated with * (*:p < 0.05, **:p < 0.01, ***:p < 0.001; p-values adjusted with the Bonferroni correction).
Concerning the walking condition, both PPG sensors estimated meanRR with no bias with respect to ECG. Nevertheless, the limits of agreement highlighted a wide range of variability among measurements (SGW5: [−78 ms, 84 ms]; shimPPG: [−80 ms, 78 ms]; see Fig. 2). Additionally, we observed a significant overestimation and broad limits of agreement for all the other HRV features considered. Particularly, the overestimation trend increased at the increase of the measurements’ magnitude.
Fig. 2. BA analysis results for the SGW5vsECG and shimPPGvsECG comparisons of meanRR (ms) estimates during walking. For each comparison we report the bias and the limits of agreement (1.96 standard deviations interval around the bias). Biases that significantly differed from 0 are indicated with * (*:p < 0.05, **:p < 0.01, ***:p < 0.001; p-values adjusted with the Bonferroni correction).
28
G. Rho et al.
2.6 Discussion In this study, we investigated the accuracy of the commercially-available SGW5 to monitor HR dynamics in a group of healthy volunteers during a resting state and a walking condition. To this aim, we estimated the PRV and its main features from the SGW5’s PPG built-in sensor, and we compared them with the HRV features obtained from the Shimmer3 ECG wearable unit [7]. Additionally, the outcome of the comparison between the SGW5-PRV features and the ShimmerECG-HRV ones was further compared with that obtained by replacing the SGW5 with a research-grade PPG sensor, i.e., the Shimmer3 GSR + unit. At resting state, our preliminary results show the SGW5 as a reliable wearable tool to provide unbiased estimates of the average HR (from PRV time series). Moreover, the correlation analysis of SGW5 HRV features against ECG highlighted significant correlation coefficients with those observed for shimPPG for relevant features such as stdRR, LF, HF, and SD2. Nevertheless, Bland Altman’s analysis showed a tendency for the SGW5 to estimate HRV features with a higher bias compared to shimPPG. Given the important informative content of these features (especially the frequency ones) linked to the autonomic nervous system, such bias could limit the reliability of the inference on sympathovagal balance. It is worthwhile noting that, while ECG is characterized by sharp R peaks, PPG has a smoother sinusoidal nature which makes peak detection intrinsically more difficult and unprecise. On the other hand, the comparison with shimPPG could be affected by the acquisition site. Indeed, previous studies reported more accurate HRV parameters extracted from the finger, with respect to the wrist [12]. In this light, we cannot exclude an effect of the sensor position on the differences observed between SGW5 and shimPPG. Concerning the walking condition, both the SGW5 and shimPPG confirmed good reliability for the estimation of the mean RR, although BA analysis indicated a significant range of variability among measurements. The PRV-related features performed poorly compared to the resting state. These results were not totally unexpected, as they are in line with previous validation studies on different PPG wearable sensors [11, 11, 13, 21]. Indeed, such devices are known to be particularly susceptible to motion artifacts [8]. Accordingly, deriving accurate PRV time series in such a context may be tricky. In noisy scenarios, several processing techniques in the frequency domain have been proposed to accurately estimate the mean HR from PPG recordings [18, 22–24]. However, to the best of our knowledge, such approaches do not provide an alternative means to estimate HRV time series and extract HRV features. Hence, our work aimed at evaluating the reliability of the PRV time series derived from the SGW5. In conclusion, our results, although preliminary, highlight the reliability of the SGW5 as an open-source, non-invasive and low-cost wearable device to monitor mean HR. On the other hand, the reliability of PRV dynamics is still an open question and deserves further investigation. Particularly, for those parameters providing a window on sympathovagal regulation such as HF power, which is a reliable parasympathetic correlate to investigate stress, anxiety, and fatigue conditions [17], future studies should evaluate whether such HRV features extracted from the SGW5 can still be used to distinguish different psychophysiological states.
Preliminary Assessment of the Samsung Galaxy Watch …
29
Acknowledgment. Funded by the European Union. Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them. The work is supported by European Union’s Horizon Europe Research and Innovation Programme under grant agreement No. 101057103 – project TOLIFE.
References 1. Samsung galaxy watch 5. https://www.samsung.com/it/watches/galaxy-watch/galaxy-wat ch5-40mm-graphite-bluetooth-sm-r900nzaaitv/. Accessed 2 Feb 2023 2. Beh, W.K., Wu, Y.H., Wu, A.Y.A.: Robust ppg-based mental workload assessment system using wearable devices. IEEE J. Biomed. Health Inf. (2021) 3. Birkhofer, A., Schmidt, G., Förstl, H.: Heart and brain–the influence of psychiatric disorders and their therapy on the heart rate variability. Fortschr. Neurol. Psychiatr. 73(4), 192–205 (2005) 4. Bishop, S.M., Ercole, A.: Multi-scale peak and trough detection optimised for periodic and quasi-periodic neuroscience data. In: Intracranial Pressure & Neuromonitoring XVI, pp. 189– 195. Springer (2018) 5. Bland, J.M., Altman, D.: Statistical methods for assessing agreement between two methods of clinical measurement. The lancet 327(8476), 307–310 (1986) 6. Boukhechba, M., Cai, L., Wu, C., Barnes, L.E.: Actippg: using deep neural networks for activity recognition from wrist-worn photoplethysmography (ppg) sensors. Smart Health 14, 100082 (2019) 7. Burns, A., et al.: Shimmer™–a wireless sensor platform for noninvasive biomedical research. IEEE Sens. J. 10(9), 1527–1534 (2010) 8. Castaneda, D., Esparza, A., Ghamari, M., Soltanpur, C., Nazeran, H.: A review on wearable photoplethysmography sensors and their potential future applications in health care. Int. J. Biosens. & Bioelectron. 4(4), 195 (2018) 9. Hartikainen, S., Lipponen, J.A., Hiltunen, P., Rissanen, T.T., Kolk, I., Tarvainen, M.P., Martikainen, T.J., Castren, M., Väliaho, E.S., Jäntti, H.: Effectiveness of the chest strap electro-cardiogram to detect atrial fibrillation. Am. J. Cardiol. 123(10), 1643–1648 (2019) 10. He, J., Ou, J., He, A., Shu, L., Liu, T., Qu, R., Xu, X., Chen, Z., Yan, Y.: A new approach for daily life blood-pressure estimation using smart watch. Biomed. Signal Process. Control 75, 103616 (2022) 11. Hernando, D., Roca, S., Sancho, J., Alesanco, Á., Bailón, R.: Validation of the apple watch for heart rate variability measurements during relax and mental stress in healthy subjects. Sensors 18(8), 2619 (2018) 12. Lee, B.G., Lee, B.L., Chung, W.Y.: Smartwatch-based driver alertness monitoring with wearable motion and physiological sensor. In: 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 6126–6129. IEEE (2015) 13. Milstein, N., Gordon, I.: Validating measures of electrodermal activity and heart rate variability derived from the empatica e4 utilized in research settings that involve interactive dyadic states. Front. Behav. Neurosci. 14, 148 (2020) 14. Morelli, D., Bartoloni, L., Colombo, M., Plans, D., Clifton, D.A.: Profiling the propagation of error from ppg to hrv features in a wearable physiological-monitoring device. Healthc. Technol. Lett. 5(2), 59–64 (2018) 15. Nardelli, M., Vanello, N., Galperti, G., Greco, A., Scilingo, E.P.: Assessing the quality of heart rate variability estimated from wrist and finger ppg: a novel approach based on cross-mapping method. Sensors 20(11), 3156 (2020)
30
G. Rho et al.
16. Paradiso, R., Loriga, G., Taccini, N.: A wearable health care system based on knitted integrated sensors. IEEE Trans. Inf. Technol. Biomed. 9(3), 337–344 (2005) 17. Roque, A.L., Valenti, V.E., Massetti, T., Da Silva, T.D., Monteiro, C.B.d.M., Oliveira, F.R., de Almeida Junior, Á.D., Lacerda, S.N.B., Pinasco, G.C., Nascimento, V.G., et al.: Chronic obstructive pulmonary disease and heart rate variability: a literature update. Int. Arch. Med. 7(1), 1–8 (2014) 18. Schaffarczyk, M., Rogers, B., Reer, R., Gronwald, T.: Validity of the polar h10 sensor for heart rate variability analysis during resting state and incremental exercise in recreational men and women. Sensors 22(17), 6536 (2022) 19. Selvaraj, N., Jaryal, A., Santhosh, J., Deepak, K.K., Anand, S.: Assessment of heart rate variability derived from finger-tip photoplethysmography as compared to electrocardiography. J. Med. Eng. Technol. 32(6), 479–484 (2008) 20. Shaffer, F., Ginsberg, J.P.: An overview of heart rate variability metrics and norms. Front. Public Health 258 (2017) 21. Sun, B., Zhang, Z.: Photoplethysmography-based heart rate monitoring using asymmetric least squares spectrum subtraction and bayesian decision theory. IEEE Sens. J. 15(12), 7161– 7168 (2015) 22. Tajrishi, F.Z., Chitsazan, M., Chitsazan, M., Shojaei, F., Gunnam, V., Chi, G.: Smartwatch for the detection of atrial fibrillation. Crit. Pathw. Cardiol. 18(4), 176–184 (2019) 23. Tarvainen, M.P., Niskanen, J.P., Lipponen, J.A., Ranta-Aho, P.O., Karjalainen, P.A.: Kubios hrv–heart rate variability analysis software. Comput. Methods Programs Biomed. 113(1), 210–220 (2014) 24. Weiler, D.T., Villajuan, S.O., Edkins, L., Cleary, S., Saleem, J.J.: Wearable heart rate monitor technology accuracy in research: a comparative study between ppg and ecg technology. In: Proceedings of the Human Factors and Ergonomics Society Annual Meeting, vol. 61, pp. 1292–1296. SAGE Publications Sage, Los Angeles, CA (2017) 25. Zhang, Y., Liu, B., Zhang, Z.: Combining ensemble empirical mode decomposition with spectrum subtraction technique for heart rate monitoring using wrist-type photoplethysmography. Biomed. Signal Process. Control 21, 119–125 (2015) 26. Zhang, Z.: Photoplethysmography-based heart rate monitoring in physical activities via joint sparse spectrum reconstruction. IEEE Trans. Biomed. Eng. 62(8), 1902–1910 (2015) 27. Zhang, Z., Pi, Z., Liu, B.: Troika: a general framework for heart rate monitoring using wristtype photoplethysmographic signals during intensive physical exercise. IEEE Trans. Biomed. Eng. 62(2), 522–531 (2014)
EEG Microstate Clustering to Evaluate Acoustic Stimulation Phase-Locked Targeting of Slow Wave Sleep Activity Filip Cerny1,2(B)
, Vaclava Piorecka1,2 , Jan Strobl1,2 , Daniela Dudysova2 Jana Koprivova2,3 , and Marek Piorecky1,2
,
1 Faculty of Biomedical Engineering, Czech Technical University in Prague, Námˇestí Sítná
3105, 272 01 Kladno, Czech Republic [email protected] 2 National Institute of Mental Health, Topolová 748, 250 67 Klecany, Czech Republic 3 Third Faculty of Medicine, Charles University in Prague, Staré Mˇesto, Czech Republic
Abstract. Closed-loop acoustic stimulation is a promising method for enhancing properties and quality of deep sleep. Additional synchronization of slow waves seems to be the result of auditory stimulation. However, the temporal parameters of typical deep sleep slow waves may be altered. We investigated this hypothesis by performing EEG microstate (MS) analysis. Further, we evaluated benefits of analysis for determining the precision of phase-locked stimulation precision. We obtained 4 MS template maps from non-rapid eye movement 3 (NREM 3) intervals of sham group’s sleep. By backfitting these templates to short data segments following stimulation or sham markers we have shown an overall dominance of frontally oriented microstate spatially related to canonical MS D in the stimulation group, which was not observed in sham group. This MS map can be linked with frontal neuron synchronization typical for low-frequency characteristic of deep sleep. Larger occurrences of this MS map may illustrate the process of sleep deepening during the stimulation. Comparing global field power (GFP) curves for stimulated subjects with polar histograms of precision we observed a relation between targeting precision for specific phases of stimulation with uniformity of MS distribution. Subjects stimulated more precisely had more uniform MS distribution in GFP’s, mostly dominated by MS 2. If confirmed by subsequent studies this property could serve as a marker of stimulation algorithm’s precision. Keywords: EEG · Microstates · Closed-loop acoustic stimulation
1 Introduction Sleep disorders are a common problem of modern society. One of the most prominent of these disorders is insomnia, affecting around 10% of the population [1]. Inability to fall asleep or frequent arousals from sleep lead to reduced cognitive performance and increased health risks in both physical and mental domains [2]. The overall lighter sleep and higher proneness to awakenings were previously demonstrated on spectral analyses © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 31–40, 2024. https://doi.org/10.1007/978-3-031-49062-0_4
32
F. Cerny et al.
of sleep recordings with apparent higher power of alpha band and lower power of delta frequencies in delta sleep [3]. Therefore, stimulation methods for deepening sleep activity were proposed. One of these methods is the closed-loop acoustic stimulation (CLAS), in which brief acoustic stimuli are played to the sleeping subject at a specific phase of deep sleep slow oscillations obtained from real-time streamed electroencephalography (EEG) data. This method was shown to elicit additional new slow waves [4], enhancing the amount of deep sleep and thus improving or restoring healthy sleep patterns [5]. The emergence of new waves subsequent to stimulation is attributed to the sound stimuli serving as a pacemaker for neural synchronization during sleep [6]. The important aspect is the phase of the slow wave in which we apply stimulation [4]. It has been shown that different positions of sound stimuli produce different responses in EEG [7]. This phenomenon could be studied with EEG microstate analysis. EEG microstates (MS) are transient states (~100 ms) of quasi-stable EEG topographies [8]. In EEG re search across fields, it has been shown that most of the signal variance can be explained by 4 microstate topographies (canonically labeled as A, B, C, D) [9]. This ideal number of microstates has also been shown to be present in sleep, although the topographies are a bit different [10, 11]. However, there is currently no study of EEG microstate dynamics on CLAS data, even though this approach could identify the impact of auditory stimulation on the natural spatial distribution of brain activity during deep sleep. Due to the CLAS impact on slow waves, the microstate temporal parameters of deep sleep may be changed. The main goal of this study was therefore to explore the spatial topography of deep sleep slow waves elicited by CLAS, utilizing the EEG microstate analysis. Further, we investigated whether microstate analysis can be used as a predictor of CLAS precision of phase targeting.
2 Methods 2.1 Dataset The dataset used for this study comes from a polysomnography (PSG) study of the influence of acoustic stimulation on insomnia patients [6]. The dataset contains 36 paired PSG recordings of 18 insomnia subjects. However, only 14 recordings (7 subjects) passed the criterion for a minimum number of 50 stimuli per night. Each of the participants underwent two recording nights of which one was with acoustic stimulation and one was a control night without the stimulation being presented. Order of nights was randomized. Measurements were performed in the sleep laboratory at the National Institute of Mental Health (NIMH). The study was approved by the local Ethics committee. The data were recorded using 9 EEG electrodes located on positions devised from standard 10/20 electrode distribution. Electrodes corresponded to the F3, F4, Fpz, P3, P4, C3, C4, M1, M2 locations. Electrooculography (EOG), electromyography (EMG) and electrocardiography (ECG) signals were also recorded for optimal artifact removal and sleep scoring. Brainscope PSG system (M&I spol. s.r.o., Czech Republic) was used for the measurement.
EEG Microstate Clustering to Evaluate Acoustic Stimulation
33
Recorded PSG data were scored by a sleep clinician and additionally reviewed by second sleep clinician. Individual sleep phases were marked in 30 s non-overlapping segments according to the American Academy of Sleep Medicine criteria [12]. 2.2 Data Preprocessing Data were originally recorded at a sample rate of 1000 Hz and were downsampled to 250 Hz prior to the analysis. Preprocessing was performed using the Fieldtrip toolbox for Matlab [13]. Steps of preprocessing included a band-pass two-way finite-impulse response (FIR) filter, of 1652 order, in frequency range 0.5 to 30.0 Hz, with additional demean and detrend filtering. Bad epochs of the data were manually rejected. Next, segments of non-rapid eye movement 3 (NREM 3) sleep were extracted based on sleep phase tags obtained from sleep scoring. Minimal length of these segments was 30 s. 2.3 Closed-Loop Acoustic Stimulation Acoustic stimulation was performed via a real-time preprocessing of the EEG signal averaged from F3 and F4 electrodes referenced to the mastoid (M1, M2) electrodes similarly to a study by Ngo et al. [4]. Implementation was realized via a script programmed in python language. Slow wave detection was set to an amplitude threshold - 80 µV for participants younger than 30 years [4] and −40 µV for older participants [14]. Played sound stimuli consisted of two consecutive 50 ms long pink noise sounds and the sound was played through headphones with the volume set subjectively by each participant prior to the start of the recording. Participants were instructed to set the volume so that they will not be awakened but can still clearly hear it. Stimulation itself was started only during the NREM 3 sleep phase which was assessed manually during recording by laboratory personnel. For averaging the waveform of response to stimuli epochs of signal 2 seconds prior and 5 seconds after the first stimulation were extracted. Participants underwent two recording nights separated by at least one week. One of the nights included acoustic stimulations and one night was without stimulation (sham condition), the order of the nights was randomized. Figure 1 depicts average waveforms for stimulation and sham (control) groups. Detection of slow waves corresponds to the minimum trough of the slow wave and the first acoustic stimuli to the zero-time point, in case of sham the zero-time point represents a tag in the data without real stimuli having been played. To obtain the precise phase of the slow wave in which the stimulation was applied we used Hilbert transformation. These phase values were computed only for the first sound stimulation for both conditions.
34
F. Cerny et al.
Fig. 1. Effect of acoustic stimuli (blue, n = 7) vs. no stimuli presented (red, n = 7). Displayed are mean signals for each group, created as the average of F3 and F4 electrodes ± standard deviation.
2.4 Microstate Analysis Microstate analysis was used to assess the main topographies of brain activity during NREM 3 and after stimulation. All analyses were performed using the microstate toolbox by Poulsen et al. [15] for Matlab’s EEGlab environment. Firstly, the template maps were obtained from grouped trials of NREM 3 stage of the sham group. This group’s sleep was not affected by CLAS and therefore it could serve as a non-disturbed control sleep. For clustering 10000 peaks of Global field power (GFP) curve were selected with minimal peak distance of 10 ms [15]. Data were normalized and an average reference was used. For segmentation of templates the k-means algorithm was used with 100 repetitions and maximum of 1000 iterations, polarity of topographies was ignored as we were analyzing spontaneous activity [16]. Optimal number of template maps were selected based on global explained variance (GEV), Krzanowski-Lai (K-L) criterion and visual evaluation. Next, the template maps were backfitted to 5 s epochs following stimulation or sham markers in the NREM 3 stage. For backfitting the polarity was again ignored as we did not consider the new waves as evoked potentials due to their long duration. 2.5 Statistical Evaluation For statistical evaluation of the results, we performed a non-parametrical Wilcoxon’s signed-rank paired test via Matlab. Evaluated parameters were occurrence, duration, GFP and coverage of each identified microstate. P-values resulting from the analysis were corrected with Bonferroni’s correction for multiple comparisons by 16 (4 microstates × 4 parameters).
EEG Microstate Clustering to Evaluate Acoustic Stimulation
35
3 Results 3.1 Microstate Analysis Microstate analysis was performed in two steps. At first the template maps were created from NREM 3 stages of recordings without sound stimuli being played. The template maps were extracted using the modified k-means clustering algorithm with parameters defined in the methods section. Selecting the just number of microstates was evaluated by visually investigating the templates and using the GEV and K-L criterion. For 5 MS templates the GEV was 72 % a K-L was lower. For the case of 4 MS templates GEV was 69 % and K-L was higher, which suggests that the optimal number is 5. However, by visually inspecting the topographies MS 5 does not illustrate brain activity but is more likely an artifact. Therefore, 4 MS maps were selected as the template for the microstate analysis, see Fig. 2.
Fig. 2. Identified microstate topographies. MS 5 in the upper part was identified as an artifact component and removed from the analysis. Templates are sorted based on the highest GEV value. Created with microstate toolbox for [15] EEGlab.
Identified microstate template maps were in the next step backfitted to trials follow ing the first stimulation. Trials spanned 5 seconds following the marker of first stimulation or sham. For backfitting also ignored the polarity of EEG data. Parameters of MS backfitting are shown in Table 1. Statistical analysis of paired data sets showed no significant difference between stimulation and sham group. However, MS 2 in the stimulation group had a larger duration and was the most prominent in the stimulation group (Fig. 3).
36
F. Cerny et al.
Fig. 3. Comparison of backfitting of MS templates to trials following acoustic stimulation or sham markers. Visualized are average trials of GFP curve for one representative subject with precisely stimulated slow-wave activity a) stimulation, b) sham. Colors represent which MS class is assigned to the part of GFP curve.
Table 1. Evaluation of microstate parameters between stimulation and sham group. No statistically significant differences were found after correction for multiple comparisons. Stimulation group Coverage (%)
Duration (ms)
Occurence (s−1 )
GFP (peaks/s)
MS 1
14.0 ± 4.4
110 ± 47
1.37 ± 0.51
1.53 ± 0.54
MS 2
56.3 ± 13.0
282 ± 12
2.20 ± 0.61
4.15 ± 2.56
MS 3
17.8 ± 10.2
106 ± 33
1.63 ± 0.54
1.49 ± 0.80
MS 4
11.9 ± 8.9
115 ± 65
1.00 ± 0.66
1.07 ± 0.49
Sham group MS 1
23.5 ± 7.7
103 ± 21
2.31 ± 0.77
1.21 ± 0.29
MS 2
41.9 ± 9.1
174 ± 56
2.54 ± 0.76
2.24 ± 1.31
MS 3
17.0 ± 6.9
101 ± 27
1.67 ± 0.47
0.87 ± 0.34
MS 4
17.6 ± 8.6
95 ± 27
1.80 ± 0.70
0.99 ± 0.41
Resulting GFP curves for individual subjects from stimulation night were further compared with polar histograms depicting the precision of stimulation algorithm. In case of precise stimulation, the polar histogram had the highest count of stimulations near 0°. This point corresponds to the peak of the cosine wave. In Fig. 4 for the case of
EEG Microstate Clustering to Evaluate Acoustic Stimulation
37
more precise stimulation MS 2 is the most prominent. For inaccurate stimulation, the stable topographies break into different MS templates.
Fig. 4. Mean GFP curves following the stimulation tag for individual subjects of stimulation group compared to polar graphs of angles of stimulation on the slow wave. 0° was the ideal stimulation angle that we were trying to stimulate.
4 Discussion We have performed EEG microstate analysis of time trials following acoustic stimulation in a CLAS experiment of insomnia patients. However, instead of focusing on clinical aspects of insomnia, we mainly tried to pilot a novel method for evaluation of technical precision of auditory stimulation during sleep.
38
F. Cerny et al.
EEG microstates can serve as a tool to visualize the dynamics of spatial topographies of the signal following the acoustic stimulation. For the analysis we selected 4 main microstates. These topographies did not precisely link to the canonical MS topographies by Lehmann et al. [9], even though it was previously shown that MS topographies in sleep are mostly conserved from wake maps [10]. However, by spatial correlation and visual evaluation we linked MS 2 and 4 to canonical microstates D and A respectively. MS maps 1 and 3 are also roughly correlated with MS B and C with a greater deviation from the canonical templates. This may be due to a low number of EEG electrodes used and by possible differences in MS architecture for insomnia patients. To compare microstate parameters across subjects we performed paired statistical tests. We found no statistically significant difference in MS parameters between stimulation and sham. Slow waves originating from acoustic stimulation were backfitted mostly to MS 2 template (Fig. 3). This dominance of MS 2 was consistent across stimulated subjects (Fig. 4). This topography was located in frontal areas of the brain. Closest electrodes to the GFP peaks for this MS (F3 and F4) also served as the slow wave detection electrodes for the stimulation algorithm. Moreover, this frontally oriented MS topography was previously linked to low-frequency synchronization in frontal areas [11]. This topography may show a deactivation of frontal brain regions, which is attributed to deep sleep, preventing the brain from waking up. Apparent dominance of MS 2 may therefore suggest deepening of sleep related to acoustic stimulation, which benefits memory consolidation [4, 17]. We hypothesized that the observed backfitting to one specific microstate class may be related to precision of stimulation application. The method applies stimulation based on a prior slow wave detected by the algorithm. Because of this, the stimulation may be applied in a different phase of the slow wave. This is depicted in circular polar his to grams visualized in Fig. 4. Comparing the GFP curves for individual subjects from stimulation group to the polar graphs we found that in case of a more precise phase targeting we can obtain a more uniform brain topography mostly related to MS map 2. It has been previously shown that stimulations applied at a different phase of slow wave produce various brain responses [18] and elicited spindles that are out of phase may not have proper effect on memory consolidation [19]. Up-slope targeted stimulations produce more spindle/beta band activity than stimulations applied in down-slope phase of the wave [7]. From these reports it seems that the phase of applied stimulation influences the effect of stimulation. This may also be reflected in MS topographical maps. Different phases of stimulation may produce slow waves distributed differently on the scalp. We see that stimulations applied in a broader interval of phase degrees were backfitted to additional topographies related mostly to MS map 1 (Fig. 4). It may be beneficial to perform MS analysis on EEG stimulated in intentionally different phases of the slow wave to explore how the MS backfitting changes with the phase of stimulation. 4.1 Limitations Main limitation of our study is a low number of subjects analyzed. This was caused by a high number of rejected data based on insufficient number of stimulations during the recording night. Another limitation is a low number of EEG channels used for record ing, which may cause uncertainties in computing MS templates. However, it has been
EEG Microstate Clustering to Evaluate Acoustic Stimulation
39
previously shown that EEG microstate topographies can be replicated with a lower number of electrodes [20]. Further research should focus on replicating the results on a broader dataset with more recording channels, also with healthy controls.
5 Conclusion We have performed EEG microstate analysis on the data of auditory stimulated subjects during CLAS experiment. We found no significant difference in MS parameters in stimulation vs. sham trials. The most prominent MS class corresponded with frontal brain areas, which may show local deactivation related to NREM 3 sleep characteristics. We showed that GFP curve with MS topographies backfitted to data following the stimulation can visualize how precisely we were able to focus the auditory stimuli to a selected phase of deep sleep slow wave and indicate that a more precise intervention leads to a more uniform brain topography with a dominating MS 2 suggesting its higher impact on sleep enhancement. Acknowledgements. This work was supported by The Czech Science Foundation (GACR) reg. no. 22- 16874S and by the Grant Agency of the Czech Technical University in Prague (SGS), reg. No. SGS22/200/OHK4/3T/17 and reg. No. SGS21/140/OHK4/2T/17.
References 1. Ohayon, M.M.: Epidemiology of insomnia: what we know and what we still need to learn. Sleep Med. Rev. 6(2), 97–111 (2002). https://doi.org/10.1053/smrv.2002.0186 2. Blackwelder, A., Hoskins, M., Huber, L.: Effect of inadequate sleep on frequent mental distress. Prev. Chronic Dis. 18, 200573 (2021). https://doi.org/10.5888/pcd18.200573 3. Svetnik, V., Snyder, E.S., Ma, J., Tao, P., Lines, C., Herring, W.J.: EEG spectral analysis of NREM sleep in a large sample of patients with insomnia and good sleepers: effects of age, sex and part of the night. J. Sleep Res. 26(1), 92–104 (2017). https://doi.org/10.1111/jsr.12448 4. Ngo, H.-V.V., Martinetz, T., Born, J., Mölle, M.: Auditory closed-loop stimulation of the sleep slow oscillation enhances memory. Neuron 78(3), 545–553 (2013). https://doi.org/10.1016/ j.neuron.2013.03.006 5. Talamini, L.M., Juan, E.: Sleep as a window to treat affective disorders. Curr. Opin. Behav. Sci. 33, 99–108 (2020). https://doi.org/10.1016/j.cobeha.2020.02.002 6. Piorecky, M., Koudelka, V., Piorecka, V., Strobl, J., Dudysova, D., Koprivova, J.: Real-time excitation of slow oscillations during deep sleep using acoustic stimulation. Sensors 21(15), 5169 (2021). https://doi.org/10.3390/s21155169 7. Cox, R., Korjoukov, I., de Boer, M., Talamini, L.M.: Sound asleep: processing and retention of slow oscillation phase-targeted stimuli. PLoS One 9(7), e101567 (2014). https://doi.org/ 10.1371/journal.pone.0101567 8. Michel, C.M., Koenig, T.: EEG microstates as a tool for studying the temporal dynamics of whole-brain neuronal networks: a review. Neuroimage 180, 577–593 (2018). https://doi.org/ 10.1016/j.neuroimage.2017.11.062 9. Lehmann, D.: Brain electric microstates and cognition: the atoms of thought. In: Machinery of the Mind, pp. 209–224. Birkhäuser, Boston, MA (1990). https://doi.org/10.1007/978-14757-1083-0_10
40
F. Cerny et al.
10. Brodbeck, V., et al.: EEG microstates of wakefulness and NREM sleep. Neuro image 62(3), 2129–2139 (2012). https://doi.org/10.1016/j.neuroimage.2012.05.060 11. Bréchet, L., Brunet, D., Perogamvros, L., Tononi, G., Michel, C.M.: EEG microstates of dreams. Sci. Rep. 10(1), 17069 (2020). https://doi.org/10.1038/s41598-020-74075-z 12. Berry, R.B., et al.: The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology and Technical Specifications, Version 2.2. American Academy of Sleep Medicine, Darien (2015) 13. Oostenveld, R., Fries, P., Maris, E., Schoffelen, J.-M.: FieldTrip: open source software for advanced analysis of MEG, EEG, and invasive electro physiological data. Comput. Intell. Neurosci. 2011, 1–9 (2011). https://doi.org/10.1155/2011/156869 14. Papalambros, N.A., et al.: Acoustic enhancement of sleep slow oscillations and concomitant memory improvement in older adults. Front. Hum. Neuro Sci. 11 (2017). https://doi.org/10. 3389/fnhum.2017.00109 15. Trier, P.A., Andreas, P., Nicolas, L., Kai, H.L.: Microstate EEGlab toolbox: an introductory guide. 289850 16. Mishra, A., Englitz, B., Cohen, M.X.: EEG microstates as a continuous phenomenon. Neuroimage 208, 116454 (2020). https://doi.org/10.1016/j.neuroimage.2019.116454 17. Ong, J.L., et al.: Effects of phase-locked acoustic stimulation during a nap on EEG spectra and declarative memory consolidation. Sleep Med. 20, 88–97 (2016). https://doi.org/10.1016/ j.sleep.2015.10.016 18. Henin, S., et al.: Closed-loop acoustic stimulation enhances sleep oscillations but not memory performance. eNeuro 6(6), p. ENEURO.0306-19.2019 (2019). https://doi.org/10.1523/ENE URO.0306-19.2019 19. Latchoumane, C.-F.V., Ngo, H.-V.V., Born, J., Shin, H.-S.: Thalamic spindles promote memory formation during sleep through triple phase-locking of cortical, thalamic, and hippocampal rhythms. Neuron 95(2), 424-435.e6 (2017). https://doi.org/10.1016/j.neuron.2017.06.025 20. Zhang, K., et al.: Reliability of EEG microstate analysis at different electrode densities during propofol-induced transitions of brain states. Neuroimage 231, 117861 (2021). https://doi.org/ 10.1016/j.neuroimage.2021.117861
Development of an Interpretable Model for Improving Differential Diagnosis in Subjects with a Left Ventricular Ejection Fraction Ranging from 40 to 55% Katerina Iscra1(B) , Miloš Ajˇcevi´c1 , Aleksandar Miladinovi´c2 , Laura Munaretto3 , Jacopo Giulio Rizzi3 , Marco Merlo3 , and Accardo Agostino1 1 Department of Engineering and Architecture, University of Trieste, Trieste, Italy
[email protected]
2 Institute for Maternal and Child Health “Burlo Garofolo”, Trieste, Italy 3 Cardiovascular Department, ASUGI and University of Trieste, Trieste, Italy
Abstract. Distinguishing between Ischemic Heart Disease (IHD) and NonIschemic Dilated Cardiomyopathy (DCM) can often be difficult without invasive coronary angiography, especially in patients with Left Ventricular Ejection Fraction (LVEF) ranging from 40 to 50%. Moreover, although it is rare, some healthy subjects (HC) can have an LVEF of about 50% and must be differentiated from IHD and DCM patients. Global longitudinal strain (GLS) and heart rate variability (HRV) analysis are efficient diagnostic tools for different cardiac conditions. The use of interpretable machine-learning methods to direct the diagnosis is also gaining popularity. Therefore, this study aimed to produce a multinomial logistic regression model based on HRV, GLS and clinical features for differential diagnosis between DCM, IHD, and HC in cases with LVEF in a range of 40–55%. The study encompassed 73 DCM, 71 IHD, and 70 HC. The model was produced by logistic regression algorithms considering the set of selected features chosen with the information gain ratio method. The results showed that the most informative features for classification between HC, DCM, and IHD were GLS, meanRR, sex, age, and LFn. The model has a moderately high classification accuracy of 73%. Finally, the developed model with its nomograms enables probabilistic interpretation of classification output between HC, DCM, and IHD, and may support the differential diagnosis in this population. Keywords: Logistic regression · Cardiac diseases · HRV · LFEV · GLS
1 Introduction Ischemic heart disease (IHD) is a subtle pathology due to its silent behavior before progressing into unstable angina, myocardial infarction, heart failure or sudden cardiac death. Clinical diagnosis is based on symptoms, particularly chest pain, electrocardiography and echocardiography; nevertheless, only invasive coronary angiography can provide a definitive diagnosis, but it is not free from complications. Dilated cardiomyopathy © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 41–48, 2024. https://doi.org/10.1007/978-3-031-49062-0_5
42
K. Iscra et al.
(DCM) is characterized by dilatation and impaired contraction of one or both ventricles, and IHD and valvular heart disease must be excluded. DCM diagnosis, particularly in the early stages, is frequently challenging and depends on advanced echocardiography (speckle tracking analysis), cardiac magnetic resonance, and genetic testing. Left ventricular ejection fraction (LVEF) is use frequently to assess heart function and predict outcomes in everyday clinical practice [1]. However, it has some inherent drawbacks. One of them is that the LVEF cut-off has yet to be firmly established, making differential diagnosis in the so-called «grey zone» (LVEF: 40–50%) difficult [2], particularly in the early asymptomatic phase. Furthermore, in some cases also healthy subjects (HC) can present LVEF values around 50%. For this reason, additional non-invasive diagnostic approaches are required to better discriminate between DCM, IHD and HC in cases with LVEF between 40% and 55%. Therefore, the identification of reliable non-invasive biomarkers for early DCM, IHD and HC differential diagnosis is still needed. Heart rate variability (HRV) and left ventricular global longitudinal strain (GLS) have been identified in recent research as promising markers for the diagnosis and prognosis of different heart diseases [3–7]. GLS is a measure for cardiac deformation analysis that may be crucial in predicting cardiovascular outcomes [8] and, compared to LVEF measurement alone, has shown higher efficacy in diagnosing total left ventricular deterioration [3]. The HRV measurement is used to evaluate cardiac autonomic activity described by the relationship between sympathetic and parasympathetic activity. The pathophysiologic alterations associated with heart diseases and the accompanying changes in HRV [9] may provide prognostic information that may not be present in GLS. The progress of machine learning models for computer-aided diagnosis based on HRV features in conjunction with other clinical data is attracting scientific interest [10]. However, the models with limited interpretability which cannot be evaluated on the level of single input features, are hardly usable in everyday clinical practice. For this reason, interpretable models that provide output information regarding a specific disease and help to evaluate the model’s plausibility are more attractive in medicine [11]. Due to their high degree of interpretability and practical usability, approaches like linear/logistic regression, classification trees, and naive Bayes models are employed in several healthcare fields, including cardiology [12, 13]. Therefore, we aimed to produce a multinomial logistic regression model based on HRV, GLS and clinical features for differential diagnosis between DCM, IHD, and HC in cases with LVEF in a range of 40–55%.
2 Methods 2.1 Study Population and Protocol The study encompassed 73 DCM (45M/28F, aged 61 ± 14 y, LVEF: 46.7 ± 4.3), 71 IHD (58M/13F, aged 74 ± 11 y, LVEF: 49.2 ± 4.7), and 70 healthy subjects (HC, 33M/37F, aged 59 ± 21 y, LVEF: 53.7 ± 1.3). We included subjects with an LVEF, obtained by the Simpson biplane method [14], in a range of 40–55% were included in the study. The DCM patients were enrolled after clinical assessment. Coronary angiography was systematically performed in patients older than 35 and with cardiovascular risk
Development of an Interpretable Model for Improving Differential
43
factors and/or without a familial history of DCM. The assessment of IHD was based on clinical, laboratory and invasive findings [15]. IHD patients did not present acute coronary syndrome in the 3 months before the Holter monitoring. Patients with known trigger factors, such as toxic insults from alcohol or drug abuse, and tachyarrhythmias were excluded from the study. The GLS has been obtained from the speckle tracking echocardiography. The measurement of GLS was performed offline using dedicated software (TomTec Arena v2.0, TomTec Imaging Systems, Unterschleißheim, Germany). The investigators visually assessed the detected endocardial border and, if necessary, manually adapted the tracing to ensure the correct tracing of the contours. The study was conducted according to the principles of the Declaration of Helsinki. All participants released their written informed consent. 2.2 Heart Rate Variability Acquisition and Processing All subjects underwent a 24 h Holter ECG recording using the ambulatory electrocardiographic recorder SpiderView (Sorin Group, Italy) with a sampling rate of 200 Hz. The RR intervals were extracted and labeled by using SyneScope analysis software (Sorin Group, Italy). The RR interval records were cut into 5 min segments without overlap. Each RR 5-min segment was included in the analysis only if the longest ectopic beats subsequence (labeled with “ectopic” by the ECG Holter) or the longest artifact subsequence does not exceed 10 s (so-called Hearth rate Total variability [4]). The segments were interpolated with cubic spline and resampled at 2 Hz. Subsequently, in each segment, linear and non-linear HRV features were extracted. In particular, the linear parameters MeanRR, SDNN, RMSSD, NN50, and pNN50 evaluating the RR variability were calculated directly from the RR sequence, whilst in the frequency domain, the absolute powers in Low (LF = 0.04–0.15 Hz) and High (HF = 0.15–0.40 Hz) frequency bands, related to the vagal and sympathetic nerve control on the heart rhythm, were estimated from the interpolated HRV signal. Moreover, the normalized low and high- frequency powers (LFn, HFn) and their ratio (LF/HF) were calculated from the latter parameters. The non-linear analysis was carried out by calculating Poincaré plot parameters (SD1, SD2) reflecting the short and long-term variability [16] and extracting Fractal Dimension (FD) and beta exponent (betaExp) [17] quantifying the complexity of the system generating the signal. 2.3 Outliers, Features Selection and Classification The dataset was preprocessed to identify the possible outliers using the covariance estimator detection method, subsequently the feature selection was performed using the information gain ratio method, and finally the logistic regression model was produced. The Covariance estimator detection method selected the outliers subjects with a proportion of outliers of 0.09 and a proportion of points included in the estimate of 1. In all, 20 subjects were excluded: 7 DCM, 11 IHD, and 2 HC. Therefore, 66 DCM (42M/24F, aged 60 ± 14y), 60 IHD (49M/11F, aged 73 ± 11y), and 68 HC (31M/37F, aged 59 ± 20 y) were considered in the study. For feature selection, we used Information Gain Ratio [18] method, which has been widely applied in machine learning tasks to improve classification performance. The cut-off of the Information Gain Ratio’s estimates of the
44
K. Iscra et al.
informative attribute was set to 0.044. The model was produced by logistic regression algorithms considering the set of selected HRV features, age, sex, and GLS. The classification accuracy (CA), the area under the curve (AUC), F1, precision, specificity, and recall of the dataset were estimated using 5 random samples with a training set size of 70% of data. Finally, the nomograms for DCM and IHD groups have been created and used to interpret and validate the obtained model. 2.4 Nomogram A nomogram is a graphical representation or diagram that allows for the calculation or estimation of various variables based on their relationships. It consists of a set of scales or axes that are logarithmically or linearly arranged, with lines or curves connecting them. Nomograms are typically used in fields such as mathematics, statistics, engineering, medicine, and other sciences. They provide a quick and visual method for classification without the need for extensive calculations or computational devices. As a result, doctors can utilize nomograms as a diagnostic aid to assist in making diagnoses. Nomograms operate on the principle of representing variables as points or lines on a diagram, enabling the determination of desired values through summation of known values. The intersection of lines or curves corresponding to known values allows for the direct reading of the desired variable from the diagram. Orange (University of Ljubljana, Ljubljana, Slovenia) open-source data visualization, machine learning and data mining Python toolkit [19] was used for implementation of logistic regression and nomogram.
3 Results The results showed that the most informative features for classification between the three groups listed in order of importance, according to obtained Information Gain Ratio coefficients were GLS, meanRR, sex, age, and LFn for the logistic regression model. The classification performance of produced logistic regression model based on selected features on the test set is reported in Table 1. Table 1. Performance measures of the logistic regression model on test set. Model
AUC
CA
F1
Precision
Specificity
Recall
Logistic Regression
0.89
0.73
0.73
0.73
0.87
0.73
The confusion matrix obtained for the logistic regression model considering the test set of data is reported in Fig. 1. The confusion matrix shows that DCM patients were classified with a precision of 76.2%, IHD patients at 63.5%, and HC subjects at 79.1%. The produced nomogram of the logistic regression model for HC, DCM and IHD groups are reported in Fig. 2. The features are listed in order of importance allowing us to select the subset of the most informative features.
Development of an Interpretable Model for Improving Differential
45
Fig. 1. Confusion matrix obtained by logistic regression model considering the test set (size of 30%) of data.
Fig. 2. Nomograms of the logistic regression model for HC (top panel), DCM (middle panel) and IHD (bottom panel) groups.
46
K. Iscra et al.
4 Discussion Distinguishing between IHD, DCM and HC can often be difficult in subjects with LVEF ranging from 40 to 50%. GLS and HRV analysis are helpful diagnostic methods for a variety of cardiac illnesses [3, 4, 7]. Simultaneously, there is an increasing interest in using interpretable machine-learning approaches to guide the diagnosis. Therefore, this study aimed to produce a logistic regression interpretable model applied for differential diagnosis between DCM, IHD, and HC with an LVEF in a range of 40–55% based on age, sex, GLS, and HRV features. Our results showed that GLS, meanRR, sex, age, and LFn are the most informative features for the differential classification of DCM, IHD, and HC. The model accuracy was 73% and the confusion matrix for the test set of data showed that the HC group was classified with the highest precision (79.1%) and the IHD group with the lowest (63.5%). This precision difference can be explained by the model’s features, which characterized better HC and DCM patients. Indeed, the created nomograms of the HC, DCM and IHD groups partially confirmed this. In the HC and DCM nomograms, the longest length of the feature GLS implies that is the most discriminatory feature for these two groups. On the other hand, the length of the features in the IHD nomogram is similar, therefore none of them appear to be more discriminant than other. The HC nomogram can be very useful for a preliminary differentiation between healthy subjects and patients. In fact, we can see that the probability of being healthy improves with a GLS value greater than 17.5, a low meanRR, and a lower age. The other two nomograms can then be used to differentiate between the DCM and IHD groups. In particular, in both nomograms, GLS, age, and meanRR are the first, second, and third features in order of importance. For the diagnosis of DCM and IHD, the model empirically determined in both cases the threshold was around 17.2 for GLS, which is in line with the threshold published in a recent study [20]. Therefore, it is worth noting that if GLS < 17.2 is present, the probability of developing one of the two disorders increases. According to Halliday et al., the probability of DCM reduces with age while the likelihood of having IHD increases [21]. The meanRR, which evaluates the RR variability, is an indicator of the irregularities of the heart rhythms. Indeed, nomograms show that as this feature increases, so does the probability of DCM or IHD disease. The outcome of the prediction is only partially influenced by sex. DCM pathology appears to impact both males and females equally. On the other hand, according to the IHD nomogram, the likelihood of belonging to this group increases for males. This finding is not in line with the epidemiology data of IHD disease, according to which women have the same incidence rates as men ten years younger [22]. What we see in our IHD nomogram is most likely attributable to the fact that men outnumber women in our IHD group. Finally, LFn is often associated with the coordinated action of the vagal and sympathetic system [23], which can change in both pathologies. In particular, Huikuri et al. found a lower low-frequency power in IHD patients with impaired left ventricular function [24], which our IHD nomogram validated. In conclusion, we produced a model with a moderately high classification accuracy. The obtained accuracy is relevant to support the diagnostic procedure because it allows clinicians to better determine the need for additional complex, invasive exams. We also demonstrated the significance of machine learning model interpretability, as nomograms
Development of an Interpretable Model for Improving Differential
47
on the level of single features can be used to assess clinical plausibility and provide more information about thresholds for disease-related parameters. Finally, our study emphasized the significance of nomograms as a tool for probabilistic classification to support clinical decision-making in the differential diagnosis of HC, DCM, and IHD.
References 1. Jung, I.H., et al.: Left ventricular global longitudinal strain as a predictor for left ventricular reverse remodeling in dilated cardiomyopathy. J. Cardiovasc. Imag. 28, 137–149 (2020) 2. Solal, A.C., Assyag, P., Clerson, P., Contre, C., Guenoun, M., Poncelet, P., Thebaut, J.F., Irina, L.: 092 “Grey Zone” of 40–50% ejection fraction in ambulatory patient with Heart Failure. Who are these patients? Lessons from the DEVENIR study. Arch. Cardiovasc. Dis. Suppl. 2:31 (2010) 3. Ferrari, F., Menegazzo, W.R.: Global longitudinal strain or measurement of ejection fraction: which method is better in stratifying patients with heart failure? Arq. Bras. Cardiol. 113, 195–196 (2019) 4. Accardo, A., et al.: Toward a diagnostic CART model for Ischemic heart disease and idiopathic dilated cardiomyopathy based on heart rate total variability. Med. Biol. Eng. Comput. 60, 2655–2663 (2022) 5. Silveri, G., Merlo, M., Restivo, L., Ajˇcevi´c, M., Sinagra, G., Accardo, A.: A big - data classification tree for decision support system in the detection of dilated cardiomyopathy using heart rate variability. Procedia Comput. Sci. 176, 2940–2948 (2020) 6. Pastore, M.C., et al.: Speckle tracking echocardiography: early predictor of diagnosis and prognosis in coronary artery disease. Biomed. Res. Int. 2021, 6685378 (2021) 7. Schroeder, J., et al.: Myocardial deformation by strain echocardiography identifies patients with acute coronary syndrome and non-diagnostic ECG presenting in a chest pain unit: a prospective study of diagnostic accuracy. Clin. Res. Cardiol. 105, 248–256 (2016) 8. Ashish, K., Faisaluddin, M., Bandyopadhyay, D., Hajra, A., Herzog, E.: Prognostic value of global longitudinal strain in heart failure subjects: a recent prototype. Int. J. Cardiol. Heart Vasc. 22, 48–49 (2018) 9. Kleiger, R.E., Miller, J.P., Bigger, J.T., Moss, A.J.: Decreased heart rate variability and its association with increased mortality after acute myocardial infarction. Am. J. Cardiol. 59, 256–262 (1987) 10. Ahmad, M.A., Eckert, C., Teredesai, A.: Interpretable machine learning in healthcare. In: Proceedings of the 2018 ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics. Association for Computing Machinery, New York, NY, USA, pp 559–560 (2018) 11. Stiglic, G., Kocbek, P., Fijacko, N., Zitnik, M., Verbert, K., Cilar, L.: Interpretability of machine learning-based prediction models in healthcare. WIREs Data Min. Knowl. Discovery 10, e1379 (2020) 12. Salman, I.: Heart attack mortality prediction: an application of machine learning methods. Turk. J. Elec. Eng. & Comp. Sci. 27, 4378–4389 (2019) 13. Melillo, P., De Luca, N., Bracale, M., Pecchia, L.: Classification tree for risk assessment in patients suffering from congestive heart failure via long-term heart rate variability. IEEE J. Biomed. Health Inform. 17, 727–733 (2013) 14. Lang, R.M., et al.: Recommendations for cardiac chamber quantification by echocardiography in adults: an update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging. J. Am. Soc. Echocardiogr. 28, 1-39.e14 (2015)
48
K. Iscra et al.
15. Heart rate variability: standards of measurement, physiological interpretation and clinical use. Task force of the European society of cardiology and the North American society of pacing and electrophysiology. Circulation 93, 1043–1065 (1996) 16. Woo, M.A., Stevenson, W.G., Moser, D.K., Trelease, R.B., Harper, R.M.: Patterns of beatto-beat heart rate variability in advanced heart failure. Am. Heart J. 123, 704–710 (1992) 17. Higuchi, T.: Approach to an irregular time series on the basis of the fractal theory. Phys. D Nonlinear Phenom. (1988) 18. Quinlan, J.R.: Induction of decision trees. Mach. Learn. 1, 81–106 (1986) 19. Demšar, J., et al.: Orange: data mining toolbox in python. J. Mach. Learn. Res. 14, 2349–2353 (2013) 20. Chang, W.-T., et al.: The predictive value of global longitudinal strain in patients with heart failure mid-range ejection fraction. J. Cardiol. 77, 509–516 (2021) 21. Halliday, B.P., et al.: Sex- and age-based differences in the natural history and outcome of dilated cardiomyopathy. Eur. J. Heart Fail. 20, 1392–1400 (2018) 22. Lloyd-Jones, D., Adams, R.J., Brown, T.M., Carnethon, M., Dai, S., De Simone, G., Ferguson, T.B., Ford, E., Furie, K., Gillespie, C., Go, A., Greenlund, K., Haase, N., Hailpern, S., Ho, P.M., Howard, V., Kissela, B., Kittner, S., Lackland, D., Lisabeth, L., Marelli, A., McDermott, M.M., Meigs, J., Mozaffarian, D., Mussolino, M., Nichol, G., Roger, V.L., Rosamond, W., Sacco, R., Sorlie, P., Stafford, R., Thom, T., Wasserthiel-Smoller, S., Wong, N.D., Wylie-Rosett, J., American Heart Association Statistics Committee and Stroke Statistics Subcommittee (2010) Executive summary: heart disease and stroke statistics--2010 update: a report from the American Heart Association. Circulation 121, 948–954 23. Huikuri, H.V., Mäkikallio, T.H.: Heart rate variability in ischemic heart disease. Auton. Neurosci. 90, 95–101 (2001)
Fractal Characteristics of Retinal Microvascular Network in Alzheimer’s Disease and Colon Cancer in Automatically Segmented Fundus Images from the UK Biobank Database Isidora Rubeži´c1(B) , Miroslav Radunovi´c1 , Dejan Babi´c2 , Tomo Popovi´c2 , and Nataša Popovi´c1 1 Department of Physiology, Faculty of Medicine, University of Montenegro, Podgorica,
Montenegro [email protected] 2 Faculty for Information Systems and Technologies, University of Donja Gorica, Podgorica, Montenegro
Abstract. Several studies have found that in people with cancer, the incidence of Alzheimer’s dementia (AD) is very low, while in people with AD the incidence of cancer is very low. These findings imply that both diseases develop through deregulation of a common pathophysiological process, which in each disease occurs in the opposite direction. In this study, we aimed to see if these differences are reflected by changes in retinal microvascular network complexity. Remodeling of retinal microvasculature has already been explored as a potential biomarker for AD. Although some hereditary forms of colon cancer (CC) are associated with retinal microvascular abnormalities, these changes have not been extensively studied. Three groups of retinal fundus images from the UK Biobank were selected, age and gender-matched: group with AD (n = 20), group with CC (n = 20), and a control group (CG, n = 20, participants without either of these two diagnoses). Blood vessels were segmented using a prediction model based on convolutional neural network. Thirty-eight (67.3%) images were successfully segmented. Boxcounting dimension (Db) - a measure of microvascular complexity, and lacunarity () - a measure of microvascular gap distribution, were calculated. Db was lower in AD compared to CC and CG combined (mean Db, AD vs. CC+CG = 1.42 vs. 1.44, p = 0.048). Mean was not different among the groups. More research is needed to improve automatic segmentation. Analysis of a larger sample that accounts for other known factors affecting microvascular complexity such as diabetes mellitus and hypertension should be done in future. Keywords: Box-counting dimension · Retinal microvascular network · Alzheimer’s disease · Colorectal cancer · Automatic blood vessel segmentation
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 49–56, 2024. https://doi.org/10.1007/978-3-031-49062-0_6
50
I. Rubeži´c et al.
1 Introduction Alzheimer’s disease (AD) is the most common cause of dementia and a significant cause of disability and dependence in older adults [1]. Colon cancer (CC) is the third most common cancer and the second leading cause of cancer-related deaths in the world [2]. AD and CC are both age-related diseases and as life expectancy is increasing globally, so does the prevalence of these diseases and they are becoming a growing public health concern [1, 2]. Scientific evidence has shown that in people with cancer, the incidence of AD is very low, while in people with AD the incidence of cancer is very low [3–7]. This interesting pattern in comorbidities implies that both of these chronic diseases result from deregulation of a common homeostatic mechanism, which for cancer and AD occur in the opposite direction. Although they share many common risk factors such as old age, sedentary lifestyle, diabetes mellitus, and cardiovascular disease, they seem to lie at the opposite ends of the same pathophysiological path, which in case of cancer results in uncontrolled cell proliferation, and in case of AD in apoptosis and degeneration of cells [8]. So far, several potential mechanisms have been described that may be responsible for such an inverse association [9]. The brain and retina share embryological origin. Therefore, the neurodegenerative changes of the brain also affect neural retinal cells resulting in retinal morphology changes. Remodeling of retinal microvascular geometry has already been explored as a potential biomarker for AD, and direct correlation between retinal vascular complexity and cognitive performance has already been demonstrated [10, 11]. Changes in retinal morphology such as congenital hypertrophy of retinal pigment epithelium (CHRPE) are known to be associated with hereditary increased risk for development of CC [12] and other tumors (Gardner and Turcot syndrome) [13, 14]. Some cases of CHRPE have focal microvascular abnormalities [15]. However, changes in retinal vasculature in CC have not been extensively studied to the best of our knowledge. In the present study, we aimed to characterize the changes of the retinal microvascular network geometry that are typical either for patients with AD or CC by using information technologies. Since the retinal vascular tree has a fractal structure and can be visualized non-invasively and easily captured by digital fundus camera [16], our goal was to compare fractal characteristics of retinal microvascular network of patients with AD, to those in patients with CC and to control group of participants (CG) who do not have either of these two diagnoses or any other form of dementia.
2 Methods 2.1 Retinal Images We used single-field color fundus photographs from the UK Biobank database. The images were captured using TOPCON 3D OCT 1000 Mk2 camera at 45° field-of-view, centered to include both optic disc and macula, and stored in PNG format with dimensions 2048 × 1536 pixels [17]. The UK Biobank is a large biomedical database of health data from half a million UK participants aged between 40 and 69 years. It was collected over a period of 4 years (from
Fractal Characteristics of Retinal Microvascular Network
51
2006 to 2010) in several assessment centers and contains various health information obtained through verbal interview, physical examination, biochemical, radiological, and genetic tests [18]. The UK Biobank prospective study was conducted in accordance with the principles of the declaration of Helsinki and is approved by the ethics committee and institutional regulatory boards. All participants gave informed consent to participate in the study [19]. Ophthalmic assessment was performed for a subset of participants between 2009 and 2010 and resulted in a collection of 68,151 color retinal images from people between 40 and 69 years of age [17]. This database provides the possibility to associate retinal morphology characteristics with diagnoses and other relevant characteristics of participants. We accessed the data in this database after receiving approval from UK Biobank and our project was registered under number 92804. 2.2 Participant Selection In order to conduct a pilot study, we selected a small subset of 60 retinal images from the UK Biobank database. These images included three groups of participants, genderand age-matched: patients with AD (n = 20), patients with diagnosis of CC (n = 20), and control group (CG, n = 20, those do not have either of these two diagnoses or any other form of dementia). Image quality was assessed by an ophthalmologist and images of bad quality were not used. Those with history of glaucoma and high myopia greater or equal to 5 diopters where also excluded, because these conditions were shown to affect the vascular network complexity [20]. 2.3 Segmentation of Retinal Vessels, Image Binarization, and Skeletonization In order to analyze the fractal geometry of the retinal microvascular network, the image of the vascular network must be first extracted from the color raw image of the retina (blood vessel segmentation). In our study, automatic blood vessel segmentation was performed using a custom-made prediction model based on convolutional neural network trained on open access STARE database of retinal fundus images [21, 22]. The STARE dataset contains images that have been captured with a TopCon TRV-50 digital fundus camera at 35° field-of-view with dimensions 700 × 605 pixels with 24 bits (8 bits per each color channel) [21]. Of the 60 images included in the study, 22 (36,7%) did not have eye images suit-able for adequate automatic blood vessel segmentation which resulted in 18 patients in the control group, 11 in AD group, and 9 patients in CC group. Since the images obtained through segmentation process were gray-scale, we subsequentially performed binarization, followed by skeletonization using the Skeletonize 2D/3D plugin in ImageJ 1.53a [23]. These steps generated black-and-white 1 pixel wide tracings of microvascular network from each retinal fundus image (see Fig. 1). 2.4 Box Counting Fractal Analyses and Lacunarity Analyses The binarized and skeletonized images were analyzed by using ImageJ 1.53a and the FracLac plug-in. Box-counting fractal dimension (Db) and lacunarity dimension ()
52
I. Rubeži´c et al.
Fig. 1. Retinal image processing - arrows pointing direction of processing from raw color image to segmented, binarized, and skeletonized image.
were calculated for each image. Fractals are self-similar patterns across different scales. Db is a measure characterizing the microvascular complexity based on their property of self-similarity. , on the other hand, is used to quantify gap distribution (the empty space between vascular branches) within the retinal image [16]. 2.5 Statistical Analyses For statistical analyses, we performed ANOVA and T-test in SOFA statistics version 1.5.3. P-values less than 0.05 were considered statistically significant.
3 Results 3.1 Automatic Segmentation of Blood Vessels The program used for automatic segmentation of blood vessels was trained on retinal images from the STARE database [21], which were captured using a different camera and different image resolution compared to the UK Biobank images. Consequently, the blood vessel segmentation process was not successful for all 60 images in our study. Out of 60 images, 38 (63,3%) were successfully segmented, 11 in AD group, 9 in CC group, and 18 in control group. The most common problems encountered in the segmentation process were choroid layer pigmentation and optic disc margins wrongly recognized as blood vessels. The automatic segmentation did not perform well in cases where retinal images contained shadows. 3.2 Age Distribution of Participants The groups were initially matched based on sex and age. However, since not all images were successfully segmented, we compared the age among groups again. The mean
Fractal Characteristics of Retinal Microvascular Network
53
age of participants was 63.84. There was no difference in mean age among the groups (Table 1). Table 1. Age, box-counting dimension and lacunarity in CG, CC and AD ANOVA CG vs. CC vs. AD p-value
CG Mean ± SD
CC Mean ± SD
AD Mean ± SD
CG+CC Mean ± SD
TTEST AD vs. CG+ CC p-value
Age
0.521
63.2 ± 3.0
64.8 ± 3.2
63.8 ± 3.8
63.7 ± 3.1
0.948
Db
0.130
1.44 ± 0.04
1.45 ± 0.03
1.41 ± 0.05
1.44 ± 0.04
0.048*
0.140
0.40 ± 0.03
0.38 ± 0.04
0.42 ± 0.06
0.39 ± 0.03
0.068
3.3 Box-Counting Dimension and Lacunarity Analyses In order to compare retinal microvascular network complexity among AD, CC, and CG, Db and were calculated for each image (Fig. 2a–c). The one-way ANOVA analysis showed that neither Db nor were different among the 3 groups (data not shown). Subsequently, the pairwise comparisons by using T-test showed that AD group had a significantly lower Db compared to the other two groups combined (p = 0.048), (Table 1). This implies that participants with AD have reduced retinal microvascular network spatial complexity compared to the healthy people and those diagnosed with colon cancer.
Fig. 2. Db is decreased in patients diagnosed with AD compared to CG and those with CC. a) Representative segmented skeletonized retinal fundus image from a patient with AD shows low vascular complexity. b) Representative segmented skeletonized retinal fundus image from a healthy person with high vascular complexity. c) Db distribution in AD vs. AC+CC.
When it comes to lacunarity, the T-test and ANOVA test did not detect any statistically significant differences among the groups.
54
I. Rubeži´c et al.
4 Discussion In this study, a custom program for automatic blood vessel segmentation of retinal images was used. Since the program was previously trained only on the STARE database, some limitations were encountered when it was applied to UK Biobank retinal images, and 36,6% of images were not segmented successfully. In contrast to the present pilot study, the future analysis will include images derived from the entire set of 68,151 color fundus images from the UK Biobank. In that case, the first step of the analysis will be an automatic color image quality assessment. The specific characteristics of each image that should be taken into account in this process were previously described [24]. Following that, in order for segmentation step to be more successful, the program should be improved and adapted to UK Biobank image resolution, size, and quality. Alternatively, deep learning approach using the entire retinal fundus image, and not just segmented blood vessels, might be developed for disease prediction. Poplin et al. used a deep learning method to predict blood pressure and other cardiovascular disease risk factors in a person based on retinal fundus images from the UK Biobank. They also generated attention maps displaying areas that the program used the most to make the prediction. Interestingly, diastolic blood pressure was predicted by areas of images not representing blood vessels [25]. The present study demonstrated that the complexity of the microvascular network expressed as Db is decreased in patients with AD in comparison to patients with CC and CG combined. The participants were matched based on age, so the influence of aging on vascular branching complexity was excluded. This fact is in agreement with previously conducted research: Cheung et al. showed that total retinal fractal dimension was decreased in AD patients compared to age-gender-race matched controls. They also hypothesized that these changes possibly reflect similar remodeling of the brain microvasculature [10]. The present study did not detect differences in Db if CC was compared to AD or CG. This could be result of low statistical power of the tests caused by a small sample size, and/or because we did not take into account comorbidities also known to affect retinal microvascular network complexity, such as hypertension and diabetes mellitus. In real life, more than one chronic disease occurs often in the same person simultaneously, which should be accounted for in the future.
5 Conclusion Fractal characteristics of retinal microvascular network change in the presence of AD. In order to develop successful method for CC and AD prediction using retinal fundus images, future research aimed at perfecting the blood vessel segmentation process and taking into account all factors that can affect the complexity of the retinal vascular network must be conducted. Acknowledgements. None.
Fractal Characteristics of Retinal Microvascular Network
55
References 1. Alzheimer’s Association: 2022 Alzheimer’s disease facts and figures. Alzheimers Dement 18(4), 700–789 (2022) 2. Morgan, E., et al.: Global burden of colorectal cancer in 2020 and 2040: incidence and mortality estimates from GLOBOCAN. Gut 72(2), 338–344 (2023) 3. Roe, C.M., Behrens, M.I., Xiong, C., Miller, J.P., Morris, J.C.: Alzheimer disease and cancer. Neurology 64(5), 895–898 (2005) 4. Roe, C.M., et al.: Cancer linked to Alzheimer disease but not vascular dementia. Neurology 74(2), 106–112 (2010) 5. Driver, J.A., Beiser, A., Au, R., Kreger, B.E., Splansky, G.L., Kurth, T., Kiel, D.P., Lu, K.P., Seshadri, S., Wolf, P.A.: Inverse association between cancer and Alzheimer’s disease: results from the Framingham Heart Study. BMJ 344 (2012) 6. Realmuto, S., et al.: Tumor diagnosis preceding Alzheimer’s disease onset: is there a link between cancer and Alzheimer’s disease? J. Alzheimers Dis. 31(1), 177–182 (2012) 7. Musicco, M., Adorni, F., Di Santo, S., Prinelli, F., Pettenati, C., Caltagirone, C., Palmer, K., Russo A.: Inverse occurrence of cancer and Alzheimer disease: a population-based incidence study. Neurology 81(4), 322–328 (2013) 8. Zabłocka, A., et al.: Inverse correlation between Alzheimer’s disease and cancer: short overview. Mol. Neurobiol. 58(12), 6335–6349 (2021) 9. Lanni, C., Masi, M., Racchi, M., Govoni, S.: Cancer and Alzheimer’s disease inverse relationship: an age-associated diverging derailment of shared pathways. Mol. Psychiatry 26(1), 280–295 (2021) 10. Cheung, C.Y., et al.: Microvascular network alterations in the retina of patients with Alzheimer’s disease. Alzheimers Dement 10(2), 135–142 (2014) 11. Cabrera DeBuc, D., Feuer, W.J., Persad, P.J., Somfai, G.M., Kostic, M., Oropesa, S., Mendoza Santiesteban, C.: Investigating vascular complexity and neurogenic alterations in sectoral regions of the retina in patients with cognitive impairment. Front. Physiol. 11 (2020) 12. Chen, C.S., et al.: Congenital hypertrophy of the retinal pigment epithelium (CHRPE) in familial colorectal cancer. Fam. Cancer 5(4), 397–404 (2006) 13. Traboulsi, E.I., et al.: Congenital hypertrophy of the retinal pigment epithelium predicts colorectal polyposis in Gardner’s syndrome. Arch. Ophthalmol. 108(4), 525–526 (1990) 14. Munden, P.M., Sobol, W.M., Weingeist, T.A.: Ocular findings in Turcot syndrome (gliomapolyposis). Ophthalmology 98(1), 111–114 (1991) 15. Touriño, R., Rodríguez-Ares, M.T., López-Valladares, M.J., Gómez-Ulla, F., Gómez-Torreiro, M., Capeans, C.: Fluorescein angiographic features of the congenital hypertrophy of the retinal pigment epithelium in the familial adenomatous polyposis. Int. Ophthalmol. 26(1–2), 59–65 (2005) 16. Popovic, N., Radunovic, M., Badnjar, J., Popovic, T.: Fractal dimension and lacunarity analysis of retinal microvascular morphology in hypertension and diabetes. Microvasc. Res. 118, 36–43 (2018) 17. UK Biobank Eye and Vision Consortium homepage, UK Biobank homepage, https://www. ukbiobank.ac.uk/. Accessed 25 Feb. 2023 18. UK Biobank homepage, https://www.ukbiobank.ac.uk/. Accessed 25 Feb. 2023 19. UK Biobank research ethics approval, https://www.ukbiobank.ac.uk/learn-more-about-ukbiobank/about-us/ethics. Accessed 25 Feb. 2023 20. Li, M., et al.: Retinal microvascular network and microcirculation assessments in high myopia. Am. J. Ophthalmol. 174, 56–67 (2017) 21. Structured analyses of the retina homepage, https://cecas.clemson.edu/~ahoover/stare/. Accessed 25 Feb. 2023
56
I. Rubeži´c et al.
22. Popovic, N., Radunovic, M., Badnjar, J., Popovic, T.: Manually segmented vascular networks from images of retina with proliferative diabetic and hypertensive retinopathy. Data Brief 18, 470–473 (2018) 23. ImageJ, National Institutes of Health, http://rsb.info.nih.gov/ij/. Accessed 25 Feb. 2023 24. Popovic, N., Vujosevic, S., Radunovi´c, M., Radunovi´c, M., Popovic, T.: TREND database: retinal images of healthy young subjects visualized by a portable digital non-mydriatic fundus camera. PLoS One 16(7) (2021) 25. Poplin, R., et al.: Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat. Biomed. Eng. 2, 158–164 (2018)
Classification of Atrial Fibrillation ECG Signals Using 2D CNN Amina Tihak(B) , Lejla Smajlovic, and Dusanka Boskovic Faculty of Electrical Engineering, University of Sarajevo, Sarajevo, Bosnia and Herzegovina [email protected]
Abstract. The paper evaluates the significance of the extracted ECG signal features to classify and diagnose Atrial Fibrillation (AF) using a 2D Convolutional Neural Network (CNN). The input to the network is a 2D scalogram calculated by applying the Continuous Wavelet Transform (CWT) on five heartbeats segments of an ECG signal. The motivation behind converting the ECG signal to a 2D image is to use deep features, including time and frequency domain features, contained in the image, rather than the raw ECG data for training purposes. The dataset used for this research was obtained from the MIT-BIH Atrial Fibrillation (MIT-BIHAF) database. In this paper, we have identified the best hyperparameters for the 2D CNN model that could successfully predict AF with an accuracy of 99.12%. The paper showed that it is possible to observe all significant changes in the signal using a scalogram in the AF classification using 2D CNN. Keywords: Atrial fibrillation · 2D convolutional neural network · CWT scalogram · AF diagnostics
1 Introduction Atrial fibrillation (AF) represents the most frequent sustained cardiac arrhythmia in adults [1]. It is associated with a five-fold increase in the risk of stroke [2], which is the second leading cause of death and the third leading cause of death and disability combined in the world [3]. AF is also associated with a three-fold increase in the risk of heart failure, two-fold increases in the risk of dementia and all-cause mortality [2], and also in quantifiable impairment in quality of life [4]. AF is the most commonly encountered cardiac rhythm disorder [5], affecting 59.695 million adults worldwide in 2019 [6], and contributing to substantial social and medical burdens [5]. AF burden is progressively increasing globally, with a nearly 1.1-fold increase in the number of AF prevalence and a roughly 1.7-fold increase in deaths over the past 30 years [6]. The latest reports from different countries presented a trend of increase in the incidence and prevalence of AF over time while mortality seems to be decreasing [7]. AF prevalence is increasing worldwide due to its greater incidence among elders and the aging of the global population [2]. AF prevalence in the U.S. is expected to increase up to 45% by 2030, to over 12 million cases, and 14–17 million cases across Europe [2]. The likelihood of developing AF is complex and influenced by biological factors, environmental factors, © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 57–65, 2024. https://doi.org/10.1007/978-3-031-49062-0_7
58
A. Tihak et al.
genetics, epigenetics, and social determinants [8]. Alarming epidemiological predictions foresee a further increase in this condition during the next few decades [1]. The aim of the paper is a classification of AF based on extracted time-frequency features of the ECG signal using 2D CNN. Extraction of time-frequency features is performed by CWT since the result can be presented as an image in the form of a scalogram. Section 2 of the paper provides an overview of the pathophysiological origins of AF and how to recognize it in the ECG signals. Section 3 describes the implementation of the CNN model for AF classification using preprocessed and 2D-transformed ECG signals. An overview of the results is given in Sect. 4, while Sect. 5 contains conclusions and a discussion about future work.
2 Motivation AF is a heart rhythm disorder characterized by rapid disorganized electrical activity originating in the atrium which is associated with increased morbidity and mortality [8]. Normal cardiac rhythm shows regular rhythm initiation in the sinus node, followed by atrial and then ventricular activation [9]. Abnormal spontaneous firing from sources other than the sinus node known as ectopic activity is absent. The mechanisms underlying AF are complex, involving increased spontaneous ectopic activation of atrial cells and impulse return through atrial tissue [9]. AF is represented by rapid, irregular, and uncoordinated atrial contraction. During chaotic electrical impulses, the musculature of the atrium is activated and mechanical dysfunction of the atrium occurs. With the nonsynchronous contraction of the atria and ventricles, as a result, there is a retention of blood in the atrial cavity which consequently leads to the formation of blood clots, which can lead to serious complications such as stroke or death. An accurate understanding of the current epidemiology of AF is crucial to improve AF prevention measures. The clinical diagnosis of AF is based on the ECG surface, and because of the disorganized electrical activity, AF is characterized by the absence of a P wave [10]. Due to the impulse return through atrial tissue, the atrium continuously contracts causing very rapid, continuous, irregular, and chaotic activity that can be recognized in ECG signals as fibrillatory waves activity. However, because the amplitude of the P wave is relatively low, making its detection difficult, the heartbeat variability, which reflects the ventricular interbeat, was proposed as a significant biomarker for AF detection [10]. Machine learning algorithms developed for AF detection are mainly based on the mentioned heartbeat variability, while there are few of them that are based on the morphological features of the ECG signal such as the absence of P-wave or the presence of a fibrillatory wave. Compared to traditional machine learning algorithms that rely on manually extracted features, deep learning classification models can directly learn features from huge data sets. Classification of AF should include different ECG signal features and researchers are mainly focused on selecting important features and appropriate forms of presenting them to classifiers. In the case of deep learning classifiers, different forms of visual presentation of signals and their features are evaluated. In the paper, [11] authors in 2019 presented a novel approach for ECG heart-beat classification using sequences of single beats transformed into the form of 2D scalogram as an input for machine learning
Classification of Atrial Fibrillation ECG Signals Using 2D CNN
59
and deep learning classification models. The authors in paper [12] presented a CNN approach for automatic detection of AF episodes as short as 10 beats based on electrocardiomatrix (ECM) images using signals from MIT-BIH-NSR, MIT-BIH-LTAF and MIT-BIH-AF databases. In the [13] authors study three different approaches of using a 2D representation of ECG signals as an input to CNN. Spectrogram, scalogram, and attractor reconstruction within 5s windows of raw ECG signal were used for AF detection. MIT-BIH-AF was used to evaluate the performance of the proposed approaches where the scalogram achieved a little bit superior results. The study described in [14] is based on continuous wavelet transform and 2D CNNs to detect AF episodes. The authors used CWT with the Daubechies 5 mother wavelet to transform five beats of the ECG signal into a time-frequency representation of ECG signals in the form of scalograms. Since this approach demonstrated superior results, in our research we have also selected CWT scalogram as input to the 2D CNN classifier and focused on appropriate signal preprocessing and segmentation to improve the results. Also, we have explored the benefits of selecting different wavelets for the CWT focusing on P-wave time-frequency analysis [15].
3 Implementation In this paper, MIT-BIH-AF [16] a public database available at PhysioNet [17], was used. This database contains 25 long-term two-lead ECG recordings sampled at 250 Hz and 12-bit resolution, and each of the 25 signals is a little more than 10 h long [18]. Recordings were obtained from 25 different subjects. All patients in this database suffered at least one AF event during the recording [19]. The original ECG data were recorded at the Beth Israel Hospital in Boston using ambulatory ECG recorders with a typical recording bandwidth of approximately 0.1 to 40 Hz. In this study, the ECG data from this database were used as an independent validation set for the AF class. In the currently available signal database, the nine records from the database have been excluded due to missing data and information, and 16 recordings were left for observation. Heart rhythm annotations are available in the database, and the annotations are manually archived in documents denoted with the.atr suffix. Annotations refer to the heart rhythms: atrial fibrillation (AFIB), atrial flutter (AFL), junctional rhythm (J), and other types of heart rhythms (N) [18]. In the available database, 605 episodes are annotated (291 for AFIB, 14 for AFL, 12 for J, and 288 for N). The implementation process can be divided into three stages: (1) signal preprocessing, (2) applying CWT to obtain scalogram, and (3) classification as it is shown in Fig. 1. The first stage is signal preprocessing: (a) noise filtering, (b) normalization, (c) resampling, and (d) segmentation. 3.1 Preprocessing Raw ECG data is not applicable to be used due to the noise presence. The resources of noises are power supply (50 or 60 Hz), muscle contraction due to breathing (0.03 Hz), and other muscle activities (1–500 Hz) or electrode movements. Signal denoising is achievable with Finite Impulse Response (FIR) filter, however, DWT with wavelet is
60
A. Tihak et al.
Fig. 1: Flow diagram of implementation process
more suitable for ECG signal denoising due to better signal reconstruction when it comes to low-frequency signal components [20]. After signal denoising, it is needed to apply normalization and resampling for the purpose to facilitate later AF classification. The amplitude of the ECG signal is scaled from 0 to 1. Resampling of the signal is done with a frequency of 125 Hz. Segmentation of ECG data into heartbeats requires information about the peak positions of the morphological characteristic of ECG signals. Data segmentation in the time window series respectful to heart activity is based on R-peak locations, which are important for the extraction of RR intervals and afterward heartbeats. Heartbeats are represented by time series comprising a segment from the previous RR interval and a segment from the current RR interval, so the information about one-cycle heart activity includes the P wave, QRS complex, and T wave. One heartbeat time series contains 187 samples corresponding to 1.5 s (with a sampling frequency of 125 Hz) which encompass the whole cardiac cycle even for a heart rate of 40 BPM. A heart rate of 40 BPM is assumed to be the lower limit of the heart rhythm, and therefore 1.5 s is the maximum duration of a heart cycle. If the number of samples for a heartbeat is smaller than 187, zeros are padded. After described segmentation of one heartbeat, five successive heartbeats are combined and so the input ECG segment will contain sufficient information about atrial and ventricular activities. For the aim of AF binary classification, all annotations with AF are considered as “1”, and the other annotations as “0”. 3.2 CWT Scalogram of ECG Segment CWT is used for time-frequency signal representation and it is applicable for nonstationary signals like ECG. A non-stationary signal refers to a signal whose frequency characteristics change over time. CWT enables localization of time and frequency. The difference between CWT and DWT is manifested in the improvement of the deficiency in the rough representation of instability when it comes to classification. Mexican Hat wavelet is chosen as the wavelet for CWT implementation. Coefficients obtained from CWT are input to the neural network in the form of a scalogram. Mexican hat wavelet
Classification of Atrial Fibrillation ECG Signals Using 2D CNN
61
is mathematically described as: −t 2 2 2 (1 − t 2 ) ψ(t) = √ √ e 34π
(1)
Wavelet transformation is a method that maps the signal into a plane of time and scales. Each scale on the scalogram represents a certain frequency range of the timefrequency plane. During the implementation, 100 scales were used based on the Mexican Hat wavelet center frequency fc of 0.25 Hz and pre-processed ECG signal sampling frequency of 125 Hz. Segments of five heartbeats are transformed into sequences of five corresponding two-dimensional CWT patterns, which can be perceived as a 3D time-frequency representation of the ECG signal. The obtained format can be useful while differentiating AF due to a different distribution of frequencies in time. In Fig. 2, examples of a heartbeat with and without detected AF are shown. In normal conditions, the P-wave is clearly defined. In AF episodes, the P-wave is not noticeable, but f-waves can be noticed with a smaller amplitude and a higher frequency as it was mentioned earlier. Scalogram is a visual representation of the wavelet transform coefficients. Scalograms are three-dimensional graphs most often presented in the 2D form of a heat map, where the vertical axis represents scales, the horizontal axis represents time, and the amplitude of the CWT coefficient to a certain scale at some point in time is represented by color. Scales can be converted into frequencies when the sampling frequency is known, as it is shown in the Fig. 2.
Fig. 2: (a) (b) 1D time domain signal in normal and AF conditions. (c) (d) 2D frequency-time plane based on CWT in normal and AF conditions.
62
A. Tihak et al.
For the purpose of AF classification equal size is chosen for sets containing AF and normal beats, and overall 10 000 segments were used. Segments were randomly chosen and the sets were balanced by applying undersampling and oversampling. Training and test sets were extracted with the proportion 4:1. Test set was used as a validation set, too. 3.3 CNN Architecture CNN is a specific type of neural network suitable for processing data such as images and time series. CNN models have been widely applied in the automatic learning and classification of diseases (i.e. automatic classification of skin lesions, COVID X-ray classification). It has been shown that CNN has good detection performance in the diagnosis of diseases using biosignals such as ECG and EEG [21]. CNN comprises convolutional, pooling, and fully-connected layers. CNN can be used for image classification because the high dimensionality of images is reduced without information loss in the convolutional layer. Convolutional layers Every convolutional block has its own layers and on each layer, filters are placed. A convolution operation is performed on the input data to create local connections so that each input region is connected to a unit in the next layer. By applying convolution operation to the input map of the image, a two-dimensional activation map is obtained, and that is the response of the filter at each position. The goal of the process is to train the neural network to activate the filter in cases where a specific feature is needed to be recognized. Several types of filters are used for every layer, and the output of the layer is a 2D matrix, where the output depths are represented. The created network contains two convolutional blocks with two layers each. Scalogram images represent the input to the first convolutional layer. The dimensions of the outputs of convolutional layers are defined by the size of the feature map, and by the number of convolutional kernels. Pooling layers A pooling layer is used to reduce the resolution of the map and speed up the calculation process with spatial invariance increase (i.e. insensitivity to small shifts of features in the neural network sample). A small portion of the sample is taken and grouped to be processed so a single value is given as a result. The most commonly used functions are max pooling, mean pooling and sum pooling. Max pooling as an outcome gives the maximum value over a part of the input data, that is covered by the kernel. The number of layer components before and after the pooling process remains unchanged. The convolution and pooling layers can be repeated multiple times before applying a fully-connected layer and thus providing output suitable for classification tasks. Two max-pooling layers are included in the CNN model, and their input is the output of the preceding convolutional layer. After the operation of convolution is applied, the maxpooling layer is used to decrease the size of the input received from the previous layer. The values of x and y are two times downsized. The outputs of the max-pooling layers are the inputs to the dropout layer. The dropout layer does not change the dimension of the map, its function is to reduce the chance of oversampling by randomized training. Fully-connected layer A fully-connected layer has a role of a classifier, which classifies input images into different classes as defined in a training dataset. Flexible non-linear combinations can be performed in a fully-connected layer so the features are integrated and the best possible classification is provided. Two fully-connected layers are included
Classification of Atrial Fibrillation ECG Signals Using 2D CNN
63
in the CNN model (see Fig. 3). The input of the first fully-connected layer is the 1D output of the flatten layer, which comes after the second max-pooling and dropout layer. The size of the fully-connected layer’s output is the same as the number of neurons in the layer. After the first fully-connected layer, the dropout layer and the second fully-connected layer are added.
Fig. 3: Schematic illustration of the CNN model
4 Results In the paper performance evaluation of the proposed 2D CNN classifier model was performed using threshold and ranking metrics. First, the threshold evaluation metrics were used to evaluate the generalization ability of the trained classifier. In this case, the evaluation metrics are used to measure and summarize the quality of the trained classifier when tested with unseen data [22]. Accuracy is one of the most common metrics used in practice by many researchers to evaluate the generalization ability of classifiers. But the other ones are precision, sensitivity, specificity, and F1-score. Evaluation of the classifier model resulted in 99.12% accuracy, 98.62% precision, 99.59% sensitivity, 98.66% specificity and an F1-score value of 0.991. Second, the ranking metrics were employed as an evaluator for classification model selection. In this case, the evaluation metric task is to determine the best classifier among different types of trained classifiers that focus on the best future performance when tested with unseen data [22]. In the paper accuracy, and loss graphs were used as ranking graphical-based metrics.
Fig. 4: (a) Evaluated accuracy vs. validation accuracy. (b) Evaluated loss vs. validation loss.
64
A. Tihak et al.
The trends in Fig. 4 evaluate the accuracy and loss values in the training and validation phase of classification through the epochs. The horizontal axis shows the graph iterations and the vertical axis represents the achieved accuracy and loss values. After 38 epochs proposed classification model achieved an accuracy of 0.9981 and a loss of 0.033.
5 Conclusion and Discussion In the paper, we have presented that the CWT is an adequate tool for extracting the features of the ECG signal necessary for detection of AF using the 2D CNN. The selected approach with specific signal preprocessing demonstrated that the majority of significant changes in the ECG signal were clearly represented as time-frequency features contributing to classification results. The performance of the CNN model was evaluated, and results showed 99.12% accuracy, 98.62% precision, 99.59% sensitivity, 98.66% specificity, and an F1-score value of 0.991. The loss value and AUC are 0.033 and 0.9981 respectively. Obtained results proved the superiority of CNN among other existing algorithms and CWT preserved features needed for appropriate AF classification. Future work should address balancing the model performance and preprocessing efforts. The implemented approach should be evaluated using more ECG signals and combining signals from more databases to avoid possible bias and improve generalization.
References 1. Saglietto, A., Ballatore, A., Xhakupi, H., De Ferrari, G.M., Anselmino, M.: Atrial Fibrillation and dementia: epidemiological insights on an undervalued association. Medicina 58(3), 361 (2022) 2. Rowan, C.J., Seabright, M.A., Rodriguez, D.E., Linares, E.C., Gutierrez, R.Q., Adrian, J.C., Cummings, D., Beheim, B., Tolstrup, K., Achrekar, A.: Very low prevalence and incidence of atrial fibrillation among Bolivian forager-farmers. Ann. Glob. Health 87(1) (2021) 3. Feigin, V.L., et al.: World Stroke Organization (WSO): global stroke fact sheet 2022. IJS 17(1), 18–29 (2022) 4. Wu, J., et al.: Temporal trends and patterns in atrial fibrillation incidence: a population-based study of 3· 4 million individuals. Lancet Reg. Health Eur. 17, 100386 (2022) 5. Essien, U.R., Kornej, J., Johnson, A.E., Schulson, L.B., Benjamin, E.J., Magnani, J.W.: Social determinants of atrial fibrillation. Nat. Rev. Cardiol. 18(11), 763–773 (2021) 6. Li, H., et al.: Global, regional, and national burden of disease study of atrial fibrillation/flutter, 1990–2019: results from a global burden of disease study. BMC Public Health 22(1), 1–13 (2022) 7. Shiyovich, A., et al.: Sex-specific contemporary trends in incidence, prevalence and survival of patients with non-valvular atrial fibrillation: a long-term real-world data analysis. PLoS One 16(2), e0247097 (2021) 8. DeLago, A.J., Essa, M., Ghajar, A., Hammond-Haley, M., Parvez, A., Nawaz, I., Shalhoub, J., D.C. Marshall, S. Nazarian, H. Calkins, Salciccioli, J.D.: Incidence and mortality trends of atrial fibrillation/atrial flutter in the United States 1990 to 2017. Am. J. Card. 148, 78–83 (2021) 9. Wakili, R., Voigt, N., Kaäb, S., Dobrev, D., Nattel, S.: Recent advances in the molecular pathophysiology of atrial fibrillation. J. Clin. Investig. 121(8), 2955–2968 (2011)
Classification of Atrial Fibrillation ECG Signals Using 2D CNN
65
10. Duan, J., Wang, Q., Zhang, B., Liu, C., Li, C., Wang, L.: Accurate detection of atrial fibrillation events with RR intervals from ECG signals. PLoS One 17(8), e0271596 (2022) 11. Izmozherova, I.B., Smirnovb, A.A.: ECG heartbeat classification using convolutional neural networks and wavelet transform. AIP Conf. Proc. 2174(1) (2019) 12. Salinas-Martínez, R., De Bie, J., Marzocchi, N., Sandberg, F.: Automatic detection of atrial fibrillation using electrocardiomatrix and convolutional neural network. CinC 1–4 (2020) 13. Krol-Jozaga, B.: Atrial fibrillation detection using convolutional neural networks on 2dimensional representation of ECG signal. Biomed. Signal Process. Control 74, 103470 (2022) 14. He, R., Wang, K., Zhao, N., Liu, Y., Yuan, Y., Li, Q., Zhang, H.: Automatic detection of atrial fibrillation based on continuous wavelet transform and 2D convolutional neural networks. Front. Physiol. 9, 1206 (2018) 15. Diery, A., Rowlands, D., Cutmore, T.R., James, D.: Automated ECG diagnostic P-wave analysis using wavelets. Comput. Meth. Prog. Bio. 101(1), 33–43 (2011) 16. Moody, G.: A new method for detecting atrial fibrillation using RR intervals. Proc. Comput. Cardiol. 10, 227–230 (1983) 17. Goldberger, A.L., et al.: PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 10(23), e215–e220 (2000) 18. Zhang, H., Dong, Z., Gao, J., Lu, P., Wang, Z.: Automatic screening method for atrial fibrillation based on lossy compression of the electrocardiogram signal. Physiol. Meas. 41(7), 075005 (2020) 19. Keidar, N., Elul, Y., Schuster, A., Yaniv, Y.: Visualizing and quantifying irregular heart rate irregularities to identify atrial fibrillation events. Front. Physiol. 12, 637680 (2021) 20. Madan, R., Singh, S.K., Jain, N.: Signal filtering using discrete wavelet transform. IJRTER 2(3), 96 (2009) 21. Xiao, C., Sun, J.: Introduction to deep learning for healthcare (2021) 22. Hossin, M., Sulaiman, M.N.: A review on evaluation metrics for data classification evaluations. Int. J. Data Min. Knowl. Manag. Process. 5(2), 1 (2015)
Feature Selection for Arrhythmia Classification Using Statistical Tests Amina Tihak(B) , Amna Grahic, and Dusanka Boskovic Faculty of Electrical Engineering, University of Sarajevo, Sarajevo, Bosnia and Herzegovina [email protected]
Abstract. In this paper, we evaluate the statistical significance of features enabling us to differentiate between the signals obtained from healthy patients and patients with some type of cardiac arrhythmia. The aim of our research is to obtain a unique feature subset from an original multi-domain feature set according to a filtering-based selection method, which selects the relevant features where the redundant and irrelevant features are removed. Feature selection was implemented us ing a statistical test appropriate for the feature distribution. When the normality assumption is satisfied, an unpaired t-test was performed, and an otherwise nonparametric Wilcoxon–Mann–Whitney test. Statistical based feature selection was performed by comparing ECG signals from the MIT-BIH Normal Sinus Rhythm (NSR) and MIT-BIH Arrhythmia (AR) Database. Keywords: Cardiac arrhythmias · Statistical significance · Feature selection · Unpaired t-test · Wilcoxon–Mann–Whitney test
1 Introduction Cardiac arrhythmias are cardiac rhythm disorders that comprise an important epidemiological and public health problem [1]. As per the data available, 15% of the world population is suffering from some type of arrhythmia [2]. Cardiac arrhythmias are relatively common in the general population and their prevalence increases with age [3]. Thus, the incidence and prevalence of any type of arrhythmias may be actually higher than generally considered [4]. They occur more frequently in elderly persons, people with a long history of smoking, patients with underlying ischemic heart disease, and patients taking certain drugs or having various systemic diseases [3]. Arrhythmias are abnormal heart rhythms that can be life-threatening [5]. Among the numerous existing cardiac arrhythmia classifications [6], cardiac arrhythmias can roughly be divided into four common types of arrhythmia: extra beats, supraventricular arrhythmias (SVA), ventricular arrhythmias (VA), and bradyarrhythmias [7]. The vast majority of supraventricular arrhythmias are not imminently lethal [6], but atrial fibrillation (AF), the most common sustained cardiac arrhythmia [8] in clinical practice, is associated with an increased risk of developing stroke, systemic embolism, and heart failure [9]. Other most common and life-threatening forms of SVA are atrial flutter (AFL), and paroxysmal supraventricular tachycardia (PSVT) [10]. On the other hand, © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 66–76, 2024. https://doi.org/10.1007/978-3-031-49062-0_8
Feature Selection for Arrhythmia Classification
67
ventricular arrhythmia requires immediate diagnosis and preventive therapies since it may present with hemodynamic instability and/or degenerate into ventricular fibrillation (VF) and sudden cardiac death (SCD) [6]. Although ventricular arrhythmias are comparably less common, they are nonetheless associated with significant morbidity and a higher risk of mortality [11]. The other less prevalent and potentially non-lethal forms of arrhythmias include bradyarrhythmias and extra beats such as premature atrial contractions, premature ventricular con tractions (PVC), extrasystoles, and premature junctional contractions [10]. Feature extraction and feature selection are the most intensive points for cardiac arrhythmia detection [12]. Feature extraction is the process of extracting the most distinct features present in a dataset which are used to represent and describe the data [13]. Feature selection is referred to the process of obtaining a subset from an original feature set according to a certain feature selection criterion, which selects the relevant features of the dataset by removing irrelevant and redundant features [14]. Both feature extraction and feature selection have the advantages of improving learning performance, increasing computational efficiency, decreasing memory storage requirements, and building better generalization models [15]. Moreover, the selection enhances the comprehensibility of data and facilitates better visualization of data [14]. The performance of the feature extraction and selection method is usually evaluated by the machine learning model. The commonly used machine learning models include Naive Bayes (NB), K-means, KNearest Neighbor (KNN), Random forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), etc. In recent decades, machine learning models have been actively used for arrhythmia classification, resulting in significant performance improvements but the challenges remain in their practical application. In the paper, feature extraction was performed on ECG signals with the aim of providing a dataset with features that may provide useful information for cardiac arrhythmia detection [16]. After that, feature selection was performed because it is very important to identify the most significant features related to the disease in medical diagnosis. Appropriate feature selection based on statistical tests was performed to differentiate arrhythmia and normal sinus rhythm electrocardiogram (ECG) signals. Section 2 of the paper provides motivation, including an overview of the pathophysiological origin of cardiac arrhythmias which are caused by an aberrant electrical impulse origination and/or propagation; and a short literature review. In Sect. 3, the evaluation of the statistical significance of the features used to differentiate the signals corresponding to normal sinus rhythm and cardiac arrhythmias is described. The implementation comprises ECG signal preprocessing, features extraction; selecting an appropriate statistical test, and calculating statistical significance. An overview of the results is given in Sect. 4, while Sect. 5 provides conclusions and discussion about future work.
2 Motivation Arrhythmia occurs when there is a fault in the electrical activity in the heart muscle, causing the heart to beat irregularly and in an uncoordinated way [17]. In recent years there have been significant improvements in understanding the pathophysiological mechanisms underlying the development of cardiac arrhythmias. Cardiac arrhythmias may
68
A. Tihak et al.
occur as a result of enhanced or abnormal im pulse formation (i.e. focal activity), conduction disturbances (i.e. reentry), or abnormal initiation or propagation of electrical excitation signals within the heart that leads to cardiac arrest [18]. Recent studies indicate that inflammation, fibrosis, and even autoimmune mechanisms could facilitate the development of arrhythmias [6]. With a better understanding of the pathophysiological mechanisms underlying the development of cardiac arrhythmias, there has been increasing interest in preventative measures to reduce their prevalence [11]. Recognizing episodes of cardiac arrhythmias in ECG signal automatically is an essential task for diagnosing the abnormalities of cardiac muscle [19]. ECG is a noninvasive diagnostic tool that evaluates changes in the electrical activity of the heart muscle over time [16]. Cardiac arrhythmias are detected based on the morphological characteristics of the ECG signal. However, the ECG signal is nonlinear with a low amplitude, leading cardiologists to neglect small changes. In addition, the manual diagnosis of ECG signals is time-consuming and cumber some because it is recorded over long periods [16]. To overcome these problems, many machine learning algorithms have been proposed to improve the accuracy of cardiac arrhythmias diagnosis. Heart rate variability (HRV), contributed to enhancing the diagnostic value brought by the clinical diagnostic techniques, by providing additional information and data that can support clinical decision-making and lead to standardized, reliable, and early diagnoses [20]. The foundations of the research linked to the HRV phenomena were established in [21] providing standardized nomenclature and definitions of terms and specification of standard methods of measurement. HRV analysis through machine learning is creating a major impact in research and the world at large, making it possible to accurately antedate diseases, lower healthcare costs, and help patients make the right decision, with regard to treatments and therapies [22]. For the last ten years, we have witnessed a significant interest in this topic and the authors addressed the issue of the large number of HRV features that could be used. The effectiveness and efficiency of machine learning solutions depend on the feature selection process with the objective to use only relevant features. The authors in [23] emphasize the importance of the selection of HRV time-domain, frequencydomain, and non-linear features to assess progress in clinical interventions, and that the selection process can be guided by appropriate studies and supplemented by values from specialized populations. They have recognized that the specific context of HRV analyses applications demands a specific approach. A comprehensive list of 70 HRV features was identified in [24], and afterwards, the features were validated to test if they could be used to differentiate between resting and at least one other physiological state [25]. The authors selected thirty-one features assessing the statistical significance of differences between the signals corresponding to different physiological states. In another, more recent example of research, authors addressed the prediction of Sudden Cardiac Death (SCD) [26, 27]. The authors identified and reviewed more than fifty different HRV features and presented a comparative study on the prediction performance of these features. The authors applied the t-test method for feature ranking, and analyzed the difference in the performance of using only time-domain features compared to using only frequency domain features, and also the effects of a longer and shorter analysis window on the performance. The aim of this paper is to select unique HRV features set
Feature Selection for Arrhythmia Classification
69
in the context of differentiating ECG signals with cardio arrhythmias from normal sinus rhythm ECG signals.
3 Implementation The sources of the ECG records used in this paper are the MIT-BIH NSR Database [28] and MIT-BIH AR Database [28, 29]. The MIT-BIH NSR Database contains 18 longterm ECG signals sampled at 128 Hz. The duration of each recording is about 24 h [30]. Signals were obtained from 18 subjects, 5 men aged from 26 to 45 years and 13 women aged from 20 to 50 years. Subjects included in this database were found to have had no significant arrhythmias. Each recording includes two types of reference annotation made by experts: rhythm annotations and beat annotations [30]. The MIT-BIH AR Database contains 48 signals sampled at 360 Hz, and each of the 48 signals is a little more than 30 min long [31]. Signals were obtained from 47 subjects, 25 men aged from 32 to 89 years and 22 women aged from 23 to 89 years. The ECG signals were extracted from two leads with 11-bit resolution over a 10 mV range and band pass-filtered at 0.1–100 Hz [7]. Twenty three recordings were selected randomly from a set of 4000 24-h ambulatory ECG recordings collected from patients at Boston’s Beth Israel Hospital; the remaining 25 recordings were selected from the same set to include other clinically significant arrhythmias that may not be well-represented in a small random sample group because they are less common [31]. Both rhythm and beat annotations are available also for this dataset [30]. The ANSI/AAMI EC57:1998 standard recommends grouping the heartbeat medical annotations into five classes. The five classes should be grouped into five main beat classes. The five classes are nonectopic beats (N), fusion beats (F), supraventricular ectopic beats (SVEB), ventricular ectopic beats (VEB), and unknown beats (Q) [32]. Both databases contain recordings that have been excluded due to errors or an insufficient number of samples to calculate certain features. The implementation process can be divided into three stages: (1) signal preprocessing, (2) feature extraction, and (3) feature selection as it is shown in Fig. 1. The first stage is signal preprocessing which comprises input data segmentation and noise filtering. ECG signal is a non-stationary signal that changes its statistical characteristics over time. The purpose of the segmentation is to divide a signal into several epochs with the same statistical characteristics such as amplitude and frequency [33]. In this paper, we have used constant segmentation where the ECG recordings from both leads were segmented into ten-minute lasting segments. As a result, 78 ten-minute recordings of normal sinus rhythm, and 168 ten-minute recordings of different types of arrhythmia were used for further analysis. In continuous monitoring, ECG signal recordings are corrupted by different noises which mainly include baseline wander (BLW) and power line interference (PLI). The presence of all types of noises results in an inaccurate estimation of morphological characteristics and feature extraction. BLW is a physiological low-frequency noise that occurs because of breathing during ECG signal record ing. FIR Kaiser high-pass filter with a cutoff frequency of 2.5 Hz was used to eliminate BLW noise. PLI is a non-physiological low-frequency noise that occurs because of interference with other surrounding devices or recording equipment itself. For PLI noise elimination, FIR Kaiser low-pass filter with
70
A. Tihak et al.
Fig. 1. Implementation process
a cutoff fre quency of 10 Hz was used. Also, noise filtering is important due to the increased sensitivity of QRS complex detection. The Pan-Tompkins algorithm was used for detecting R peaks within the QRS complex. In Fig. 2 original and filtered ECG signal recordings with detected R peaks of normal sinus rhythm and arrhythmia are illustrated.
(a)
(b)
(c)
(d)
Fig. 2. Original and filtered signal of: (a) normal sinus rhythm, and (b) cardiac arrhythmia. Detected R peaks within QRS complex in signals of: (c) normal sinus rhythm, and (d) cardiac arrhythmia.
Feature Selection for Arrhythmia Classification
71
The detection of R peaks in the ECG signals enabled the extraction of the heart rate (HR) signal. Heart rate variability (HRV) is the fluctuation in the time intervals between adjacent heartbeats. These variabilities can be quantified using HRV features from different domains. Table 1 provides an overview of the multi-domain features derived from the HR signals separately for each of the groups of signals. The initial feature set comprises thirteen time domains, sixteen frequency domains, and twelve non-linear features, in total 54 features. Time domain parameters can be extracted using previously detected R peaks. From a series of direct measurements of the RR interval’s duration or from the difference in their mutual durations, it is possible to calculate complex statistical measures of the time domain. But in order to obtain frequency domain features spectral transformation of the QRS complex is required through power spectral density, which can be obtained through various spectral analysis approaches. Frequency domain features describe how power is dispersed as a function of frequency. In order to apply the Welch method on HR signal, it is necessary to interpolate and resample the signal, which can lead to the loss of information in certain frequency bands. In order to check whether mentioned information losses occur, the autoregressive method was used. Non-linear analysis methods differ from the conventional HRV methods because they do not assess the magnitude of variability but rather the quality, scaling, and correlation properties of the signals [34]. In other words, they are related to the unpredictability, fractability, and complexity of the signal [34]. The Python programming language was used in the stage of signal preprocessing and HRV feature extraction. The mean values of HRV features by the group together with the associated standard deviation are shown in Table 1. The same type of arrhythmia may have diverse morphology in different patients. Therefore, it is challenging to extract unique features for certain types of cardiac arrhythmias. The increased appreciation of the clinical potential of HRV analysis has led us to measure the significance of the differences in the HRV feature values necessary to differentiate the signals obtained from healthy patients and patients with some type of cardiac arrhythmia. In order to com pare these differences it is necessary to apply an appropriate statistical test. An assessment of the normality is a prerequisite for many statistical tests because normal distribution is an underlying assumption in parametric testing [35]. In this paper, two main methods of assessing normality were used. Although there are various methods for normality testing for small sample sizes, in this paper Shapiro–Wilk test was used as a numerical method of assessing normality. For relatively low sample sizes, as is our case, numerical methods have the disadvantage of sometimes not being sensitive enough, and because of that the Q-Q plot as a graphic method was also used for normality testing. In order to achieve a normal distribution it is necessary to perform outlier removal by using data normalization. In this paper, three different methods of data normalization were used: Z-score, IQR, and min-max normalization. By comparing the results of these methods, IQR normalization was selected as the best one, as it was the only method that showed changes in the distribution of the data. In the case of normal distribution, the parametric unpaired t-test was used. We have used an unpaired test because we have two unrelated (i.e., independent or unpaired) groups of samples. It is important to note that a two-tailed t-test was used be cause we wanted to determine whether the HRV feature values obtained from two groups of patients have a statistically
72
A. Tihak et al.
significant difference in their means. Note that, an unpaired two-sample t-test can be used only when the variances of the two groups are equal. This can be checked using F-test. If the normal distribution is not represented, regardless of the normalization of the data, it is necessary to use a non-parametric Wilcoxon–Mann–Whitney test. In the paper, a significance level value of 0.05 was used, and if the value of the p for the test is less than 0.05, the feature was considered significant, otherwise, the feature is not indicating sufficient difference for distinguishing between two groups of signals. The R programming language was used for statistical data analysis.
4 Results An overview of the statistical test results is presented in the Table 1. The results showed that there are certain statistically insignificant differences in the values of HRV features necessary to differentiate signals obtained from healthy patients and patients with some type of arrhythmia. A total of fifty-four multi-domain HRV features were obtained. It should be emphasized that the HRV frequency features were counted twice, considering that they were obtained with two different spectral analysis methods. It was mentioned earlier that if the Welch method is applied, it is necessary to interpolate and resample signals, which can lead to the loss of information in certain frequency bands. In order to check whether mentioned information losses occur, the autoregressive method was used. The mean values of extracted multi-domain HRV features by the group along with the associated standard deviation and p values of performed statistical tests are shown in Table 1. Rows in which the p-value is less than the adopted level of significance (0.05) are highlighted. This allowed us to single out only those features whose statistical difference in feature values enables the differentiation between the before-mentioned groups of signals. From the results, it can be noted that only two time-domain features showed a statistically insignificant difference in the comparison of the two groups of signals. This means that they can be considered insignificant and also can be excluded from the initial set of HRV features. The results showed that there is a loss of information in the high-frequency and low-frequency bands if the Welch method is used in regard to the autoregression method of signal spectral analysis. One-third of non-linear HRV features can be considered insignificant. The results showed that there is a certain number of HRV features that can be excluded from the initially extracted feature set and that they can be considered statistically insignificant. The remaining features are considered significant for classification. Further analysis should be conducted through the classification process.
Feature Selection for Arrhythmia Classification
73
Table 1. HRV feature values comparison. Bolded features indicate statistical significance for differentiation. HRV feature
MIT/BIH NSR Database
MIT/BIH AR Database
p-value
SDNN
93.3599 ± 39.3332
152.5086 ± 133.5159
0.0089
SDANN
39.6071 ± 38.5137
32.7555 ± 54.7170
0.0006
SDSD
62.2410 ± 34.1622
149.7863 ± 142.3956
p < 0.01
SDNNI
84.8282 ± 34.7731
146.0084 ± 129.4322
0.0018
RMSSD
83.6771 ± 53.3447
195.1976 ± 186.6504
p < 0.01
NN50
231.2564 ± 235.7613
284.7904 ± 221.2368
0.0404
pNN50
27.0366 ± 22.8072
28.4484 ± 28.3692
0.0069
NN20
419.4103 ± 213.9022
462.7964 ± 176.1792
0.0664
pNN20
50.9886 ± 22.1214
62.3888 ± 21.3940
0.0008
HRVTI
13.9485 ± 6.1590
15.2801 ± 9.8112
0.9899
TINN
575.1202 ± 243.8826
700.4585 ± 179.3983
0.0002
mean RR
0.7329 ± 0.1356
0.8276 ± 0.1801
0.0003
mean HH
184.5573 ± 94.2272
79.4660 ± 16.5965
0.0049
32482.2567 ± 246682.6582
0.2911
Time domain
Frequency domain – Welch method aVLF
2048.9824 ± 6384.1185
aLF
2330.3595 ± 8906.4741
54819.8742 ± 314796.1039
0.9529
aHF
1973.6614 ± 4992.9560
25623.5105 ± 68948.8457
p < 0.01
aTotal
6353.003 ± 20095.4120
112925.6413 ± 587315.9836
0.0017
rVLF
34.4509 ± 13.2475
20.5889 ± 18.4129
p < 0.01
rLF
35.0319 ± 10.8705
23.5872 ± 13.7981
p < 0.01
rHF
30.5172 ± 15.4169
55.8239 ± 25.2470
p < 0.01
peakVLF
0.0157 ± 0.0040
0.0202 ± 0.0090
0.0004
peakLF
0.0645 ± 0.0285
0.0825 ± 0.0426
0.0054
peakHF
0.2510 ± 0.0925
0.2470 ± 0.0842
0.8501
logVLF
6.9465 ± 0.9714
6.9138 ± 2.2775
0.2351 (continued)
5 Conclusion and Discussion Feature extraction and feature selection are the most intensive points for cardiac arrhythmia detection. Initially, features were extracted from the ECG signals of healthy patients and patients with some type of cardiac arrhythmia in three different domains. In order to select relevant features and remove irrelevant and redundant ones, feature selection was performed based on statistical tests. Appropriate statistical tests were selected after the
74
A. Tihak et al. Table 1. (continued)
HRV feature
MIT/BIH NSR Database
MIT/BIH AR Database
p-value
logLF
6.9996 ± 0.8580
7.2952 ± 2.3672
0.9529
logHF
6.7635 ± 1.1645
8.2013 ± 2.0285
p < 0.01
normLF
55.0375 ± 17.3308
33.0342 ± 21.6236
p < 0.01
normHF
44.9625 ± 17.3308
66.9658 ± 21.6236
p < 0.01
LF/HF
1.7299 ± 1.5590
0.8745 ± 2.1135
p < 0.01
Frequency domain – Autoregressive method aVLF
2333.6078 ± 79.9892
2430.0642 ± 198.1223
0.0006
aLF
4768.4708 ± 284.9397
5129.2393 ± 692.6967
p < 0.01
aHF
9290.8388 ± 886.3924
10664.5156 ± 1761.6659
p < 0.01
aTotal
16392.9175 ± 1163.2547
18223.8196 ± 2578.7501
p < 0.01
rVLF
14.2784 ± 0.7018
13.4518 ± 0.9638
0.0006
rLF
29.1307 ± 1.0867
28.2107 ± 1.2443
p < 0.01
rHF
56.5909 ± 1.7238
58.3375 ± 2.07010
p < 0.01
logVLF
7.7546 ± 0.0336
7.7926 ± 0.0771
0.0006
logLF
8.4682 ± 0.0561
8.5348 ± 0.1226
p < 0.01
logHF
9.1325 ± 0.0928
9.2616 ± 0.1607
p < 0.01
normLF
33.9939 ± 1.5081
32.6115 ± 1.7346
p < 0.01
normHF
66.0061 ± 1.5081
67.3885 ± 1.7346
p < 0.01
LF/HF
0.5158 ± 0.0345
0.4849 ± 0.0387
p < 0.01
Non-linear domain SD1
59.1681 ± 37.7205
138.0246 ± 131.9818
p < 0.01
SD2
113.8289 ± 50.7241
157.9828 ± 143.1679
0.6142
SD1/SD2
0.5653 ± 0.4061
0.9469 ± 0.4706
p < 0.01
S
23307.5934 ± 20988.8553
117529.0219 ± 241336.3995
0.0002
DFA α1
1.0076 ± 0.3074
0.6755 ± 0.2805
p < 0.01
DFA α2
0.9290 ± 0.1881
0.7177 ± 0.3350
p < 0.01
SampEn
0.9962 ± 0.3648
1.0986 ± 0.4633
0.0978
HurstExp
0.6964 ± 0.1182
0.5378 ± 0.2050
p < 0.01
dfa
0.9524 ± 0.2003
0.6811 ± 0.2936
p < 0.01 (continued)
evaluation of the normality assumption, and the observed results were considered statistically significant at a p-value of 0.05 or less. A good feature selection method should have improvement of the model performance results but less computational overhead in terms of time and dimension complexity. Our future work will compare the evaluation
Feature Selection for Arrhythmia Classification
75
Table 1. (continued) HRV feature
MIT/BIH NSR Database
MIT/BIH AR Database
p-value
CD
1.6387 ± 0.3932
1.5822 ± 0.3802
0.9606
skewRR
0.4436 ± 2.0117
0.7855 ± 3.3395
0.2217
kurtosisRR
10.8569 ± 36.3668
22.6224 ± 58.4777
p < 0.01
results of the classification approaches using the whole initial set of extracted features and using only a subset of statistically selected features for different machine learning models.
References 1. Murakoshi, N., Aonuma, K.: Epidemiology of arrhythmias and sudden cardiac death in Asia. Circ J 77(10), 2419–2431 (2013) 2. Gupta, V., Mittal, M., Mittal, V.: Chaos theory and ARTFA: emerging tools for interpreting ECG signals to diagnose cardiac arrhythmias. Wirel. Pers. Commun. 118, 3615–3646 (2021) 3. Little, J.W., Falace, D.A., Miller, C.S., Rhodus, N.L.: Little and Falace’s Dental Management of the Medically Compromised Patient, 8th edn., pp. 67–80 (2013) 4. Lazaros, G., et al.: Incidence and prevalence of cardiac arrhythmias in pericardial syndromes. RCM 23(10), 347 (2022) 5. Faust, O., Acharya, U.R.: Automated classification of five arrhythmias and nor mal sinus rhythm based on RR interval signals. Expert Syst. Appl. 181, 115031 (2021) 6. Hammerer-Lercher, A., Namdar, M., Vuilleumier, N.: Emerging biomarkers for cardiac arrhythmias. Clin. Biochem. 75, 1–6 (2020) 7. Sahoo, S., Dash, M., Behera, S., Sabut, S.: Machine learning approach to detect cardiac arrhythmias in ECG signals: a survey. Irbm 41(4), 185–194 (2020) 8. Kavousi, M.: Differences in epidemiology and risk factors for atrial fibrillation be tween women and men. Front. Cardiovasc. Med. 7, 3. Feature Selection 11 (2020) 9. Dai, H., et al.: Global, regional, and national prevalence, incidence, mortality, and risk factors for atrial fibrillation, 1990–2017: results from the global burden of disease study 2017. EHJQCCO 7(6), 574–582 (2021) 10. Albakri, A.: A meta-analysis of ECG abnormalities (Arrhythmias) in different types of heart failure. Integr. Mol. Med. 7, 1–10 (2020) 11. Patel, K.H., Hwang, T., Se Liebers, C., Ng, F.S.: Epicardial adipose tissue as a mediator of cardiac arrhythmias. Am. J. Physiol. Heart Circ. 322(2), H129-44 (2022) 12. Sahoo, S., Subudhi, A., Dash, M., Sabut, S.: Automatic classification of cardiac ar rhythmias based on hybrid features and decision tree algorithm. Int. J. Autom. Comput. 17(4), 551–561 (2020) 13. Salau, A.O., Jain, S.: Feature extraction: a survey of the types, techniques, applications. ICSC 158–164 (2019) 14. Cai, J., Luo, J., Wang, S., Yang, S.: Feature selection in machine learning: a new perspective. Neurocomputing 300, 70–79 (2018) 15. Li, J., et al.: Feature selection: a data perspective. CSUR 50(6), 1–45 (2017) 16. Kim, Y.K., Lee, M., Song, H.S., Lee, S.W.: Automatic cardiac arrhythmia classification using residual network combined with long short-term memory. IEEE Trans. Instrum. Meas. 71, 1–7 (2020)
76
A. Tihak et al.
17. European Heart Network.: Atrial fibrillation and cardiovascular diseases—a European heart network paper (2015) 18. Bukata, I.T., Tegene, E., Gobena, T., Woldesenbet, Y.M.: Prevalence and determi nants of cardiac arrhythmias and conduction anomalies in adults aged ≥ 40 years in Jimma Town, Southwest of Ethiopia: a cross-sectional study. Afr. Health Sci. 22(1), 210–219 (2022) 19. Ramkumar, M., Babu, C.G., Kumar, K.V., Hepsiba, D., Manjunathan, A., Kumar, R.S.: ECG cardiac arrhythmias classification using DWT, ICA and MLP neural networks. J. Phys. Conf. Ser. 1831(1), 012015 (2021) 20. Ponsiglione, A.M., Amato, F., Romano, M.: Multiparametric investigation of dynamics in fetal heart rate signals. Bioengineering 9(1), 8 (2021) 21. Task force of the European society of cardiology the North American society of pacing electrophysiology, heart rate variability. Circulation 93(5), 1043–1065 (1996) 22. Ishaque, S., Khan, N., Krishnan, S.: Trends in heart-rate variability signal analysis. Front. Digit. Health 3, 639444 (2021) 23. Shaffer, F., Ginsberg, J.P.: An overview of heart rate variability metrics and norms. Public Health Front. 5, 258 (2017) 24. Smith, A.L., Owen, H., Reynolds, K.J.: Heart rate variability indices for very short term (30 beat) analysis, Part 1: survey and toolbox. J. Clin. Monit. Comput. 27(5), 569–576 (2013) 25. Smith, A.L., Owen, H., Reynolds, K.J.: Heart rate variability indices for very short term (30 beat) analysis, Part 2: validation. J. Clin. Monit. Comput. 27(5), 577–585 (2013) 26. Parsi, A., O’Loughlin, D., Glavin, M., Jones, E.: Prediction of sudden cardiac death in implantable cardioverter defibrillators: a review and comparative study of heart rate variability features. IEEE Rev. Biomed. Eng. 13(5), 5–16 (2019) 27. Parsi, A., Byrne, D., Glavin, M., Jones, E.: Heart rate variability feature selection method for automated prediction of sudden cardiac death. Biomed. Signal Process. Control 65, 102310 (2021) 28. Goldberger, A.L., et al.: PhysioBank PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 101(23), e215–e220 (2020) 29. Moody, G.B., Mark, R.G.: The impact of the MIT-BIH arrhythmia database. IEEE Eng. Med. Biol. Mag. 20(3), 45–50 (2001) 30. Kleyko, D., Osipov, E., Wiklund, U.: A comprehensive study of complexity and performance of automatic detection of atrial fibrillation: classification of long ECG recordings based on the PhysioNet computing in cardiology challenge 2017. BPEX 6(2), 025010 (2020) 31. Elhaj, F.A., Salim, N., Harris, A.R., Swee, T.T., Ahmed, T.: Arrhythmia recognition and classification using combined linear and nonlinear features of ECG signals. Comput. Methods Programs Biomed. 127, 52–63 (2016) 32. Nascimento, N.M., Marinho, L.B., Peixoto, S.A., do Vale Madeiro, J.P., de Albuquerque, V.H., Filho, P.P.: Heart arrhythmia classification based on statistical moments and structural co-occurrence. CSSP 39(2), 631–50 (2020) 33. Azami, H., Mohammadi, K., Bozorgtabar, B.: An improved signal segmentation using moving average and Savitzky-Golay filter. J. Signal Process. Syst. 3, 39–44 (2012) 34. Godoy, M.F.: Nonlinear analysis of heart rate variability: a comprehensive review. J. Cardiol. Therapy 3(3) (2016) 35. Mishra, P., Pandey, C.M., Singh, U., Gupta, A., Sahu, C., Keshri, A.: Descriptive statistics and normality tests for statistical data. Ann. Card. Anaesth. 22(1), 67 (2019)
Investigating the Physiology Behind Nose Thermal Response to Stress: A Cross-Mapping Approach Federica Gioia1(B) , Mimma Nardelli1,2 , Enzo Pasquale Scilingo1,2 , and Alberto Greco1,2 1 Deparment of Information Engineering, University of Pisa, Pisa, Italy
[email protected] 2 Research Center “E. Piaggio”, University of Pisa, Pisa, Italy
Abstract. InfraRed Thermography (IRT) is an innovative monitoring tool, very beneficial in the field of psychophysiology to assess mental states in a completely non-invasive manner. IRT is contactless and it measures the temperature of the skin remotely. Thermal patterns of the skin are driven by the autonomic nervous system, thus, they reflect autonomic responses to affective stimuli. Insights into the physiology be hind changes in facial thermal patterns could improve our thermography based inferences reaching greater emotional specificity. Here, we used a cross-mapping approach to investigate the nonlinear correlations between IRT and heart rate variability (HRV), which is a powerful tool for characterizing cardiovascular activity. We used thermal signals of the nose and HRV data from 30 subjects during a resting state and under a stressful stimulation task. Our results show a statistically significant nonlinear correlation between the time series, with a significant increase in the correlation values during the stressor. Our preliminary results give a foundation to the hypothesis that the stress-induced decrease in the nose temperature is related to sympathetic induced vasomotor activity. Keywords: Thermal imaging · Nose · Heart rate variability · Cross-mapping
1 Introduction In the last decades, infrared thermography (IRT) has captured the attention of researchers in the psychophysiological field, amongst many others. Facial thermal variations have been associated with the autonomic nervous system (ANS) activity, hence allowing inferences of affective nature [1]. The ANS regulates the visceral responses, modulating the activity of its two branches, the sympathetic (SNS) and the parasympathetic systems (PNS). Many studies have explored the degree of autonomic emotion specificity, analyzing the different patterns in peripheral physiological responses to a variety of affective stimuli [2–4]. Compared to other physiological measures, this cutting-edge technology has the great ad vantage of being a contactless non-invasive diagnostic and monitoring tool. In fact, IRT measures the temperature of the skin of the subjects at a distance, by detecting the IR energy that any body, at a temperature higher than absolute zero, emits. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 77–85, 2024. https://doi.org/10.1007/978-3-031-49062-0_9
78
F. Gioia et al.
IRT has been proven to be reliable and efficient for stress monitoring in several studies [1, 5–12]. To this aim, the most relevant region of interest is the tip of the nose. The temperature of the nose decreases during stressful situations and increases due to an empathetic response [1, 10–12]. The hypothesis is that the ANS influences the vasomotor activity of the subcutaneous vessels and consequently the dynamics of the blood circulation which produces the temperature variation [12–14]. However, this is not the only process involved in temperature regulation, also sweating, respiration and metabolic processes play a crucial role [1, 12, 15]. It is now common knowledge that physiological processes have nonlinear dynamics [16]. In particular, specific aspects of the activity of human thermoregulatory system can show nonlinear dynamics that have been compared to the behaviour of theoretical nonlinear systems [17, 18]. This study aims to provide insights into the ANS correlates of nose temperature modulation in response to emotional stimuli. We want to give a foundation to the hypotheses that have prevailed in the interpretation of the results obtained in thermography studies applied to the recognition of the emotional state. Specifically, we focused on the link between the thermal signal of the nose with the heart rate variability (HRV) signal, which is a powerful tool for characterizing cardiovascular activity [19]. We used a cross-mapping approach to quantify the nonlinear correlations between IRT and HRV dynamics [20–22]. Particularly, we applied this analysis to thermal signals extracted from the nose region and to HRV time series recorded from healthy subjects during rest and under a stressful stimulus, to address the question both under parasympathetic and sympathetic nervous system predominance conditions, respectively.
2 Materials and Methods 2.1 Subjects Recruitment and Experimental Set-up Thirty healthy volunteers (20 females, age = 26.6 ± 3.6) participated in the study. None of them had any history of mental disorders or cardiovascular diseases. Coffee and smoke were forbidden on the same day of the experiment to avoid the potential vasomotor effects of these substances. The experiments were conducted in a controlled environment. In particular, the room environmental conditions were kept constant and at comfortable levels of temperature and humidity, around 24–26 z C and 40–60%, respectively. All recordings took place away from direct heat and ventilation sources. Subjects were asked to sit in a comfortable chair in the experimental room for at least 10 minutes before the acquisition started. This acclimatization period limited the thermoregulation processes during the acquisition. This experiment was approved by the “Bioethics Committee of the University of Pisa” (n. 15/2019), and each subject signed the informed consent upon arrival. The experiment comprised two main phases: a resting session and a stress stimulation, lasting 5 minutes each. We developed a Psychopy application ad-hoc for this experiment. The application ran on a laptop and guided the subject throughout the experimental phases. The protocol timeline is shown in Fig. 1. Before and after the experiment, each subject completed the State-Trait Anxiety Inventory (STAY-Y1), a psychometric test consisting of 20 self-report items on a 4-point Likert scale to measure state anxiety. Moreover, the subject was required to report its self-assessed perceived stress (PS) level
Investigating the Physiology Behind Nose Thermal Response
79
on a Likert scale from 0 (not at all) to 10 (very stressed), before and after the stressor. During the resting session, the subjects were asked to stare at a fixation cross in the middle of the screen. During the stimulation phase, stress was induced using a computerized and paced Stroop test [23]. In particular, the subject was required to solve incongruous color/semantic-meaning stimuli every 2 seconds. In addition, at the top of the screen, a counter kept track of the number of consecutive successes, as a motivational stressor. PS
PS
Acclimatization 0
5
5+x
STAY -Y1
min
10+x
STAY -Y1
Fig. 1. Protocol timeline. The dashed line indicates non-timed tasks.
We acquired the thermograms of the face of the subjects using a FLIR T640 thermal camera with a 24.6 mm lens, 640x480 pixels, NETD < 0.04 mK @ +30z and spectral range of 7.8–14 µm (Long-wave-infrared, LWIR). The sampling frequency was 5 Hz. In addition, RGB images in the visible spectrum were recorded using a Logitech HD WebCam C270. The two cameras, IR and RGB, were vertically aligned. Each subject was asked to find a comfortable position and stay still during the recordings, using a chin rest to avoid any movement artifacts. Moreover, we acquired the electrocardiogram (ECG) using a BIOPAC MP150 system, at a sampling rate of 250 Hz. Three Ag/AgCl electrodes were placed be low the right and left clavicle, and on the lower left chest to measure the ECG signal. 2.2 ECG and Thermal Time Series Processing Both ECG and thermal time series were segmented into 150-s-long segments for each experimental session (i.e., rest and stress sessions). In particular, we selected the last half of the resting session as a baseline, to make sure sub jects had completely recovered from the distress caused by the setup process. Conversely, only the first half of the stress session was used for the analysis to maximize the stressful effect and mitigate the habituation effect on the stimulation phase. Hereinafter, we will refer to these segments as Rest and Stress, respectively. The thermal time series of the tip of the nose was extracted from the thermograms. We used the RGB images to localize the region of interest (ROI), and we referred the RGB coordinates to the IR frames to obtain the thermal information. To this aim, we synchronized and registered the first IR and RGB frames. In particular, we selected the first frame of the RGB series and applied an affine-2D transformation to match the corresponding IR frame. Then, the coordinates of the centre of the ROI were automatically detected by applying the Yuval Nirkin algorithm on the registered RGB frame [24]. Furthermore, we segmented each subject’s face using a landmark-intensity-based algorithm, as explained in [25]. Thus, the number of pixels of the face of the subject was used to choose the dimension of the ROI, accounting for the inter-subjects differences in face shapes and sizes. Finally, the thermal signal was obtained as the median temperature
80
F. Gioia et al.
within the ROI for each thermogram. For each of the two sessions, we calculated the mean of the thermal signal of the nose. Concerning the ECG signals, the signals were pre-processed using the Kubios software to remove noise and artifacts and extract the Heart Rate Variability (HRV) series [26]. Specifically, interbeat interval series (RR series) were extracted from ECG signals using the Pan-Tompkins algorithm to automatically detect the QRS complexes. Then, the HRV series were extracted by interpolating the RR series through a piecewise cubic interpolation at 5 Hz. From the HRV segments related to both the Rest and Stress sessions, we extracted some of the standard features belonging to time and frequency domains, as reported in Table 1. Of note, two female subjects were excluded from the analysis due to the presence of irreducible artifacts. Table 1. Features extracted from the HRV Feature
Description
meanHRV
Mean value of HRV
stdHRV
Standard deviation of HRV
RMSSD
Root mean square of successive RR interval differences
pNN50
Percentage of successive RR intervals that differ by more than 50 ms
LF
Percentage of the total power in the low-frequency (0.04–0.15 Hz)
HF
Percentage of the total power in the law-frequency (0.15–0.40 Hz)
LF/HF ratio
Ratio of LF to HF power
2.3 Cross-mapping The cross-mapping procedure [20, 21] was applied to compute an estimate of the IR time series (Y = (Y1 , Y2 , ..., Yn )) from the HRV one (X = (X1 , X2 , ..., Xn )). We can use the delayed version of one time series to reconstruct a shadow manifold, which is an approximation to the true attractor of the dynamical system to which the series belongs. According to Takens’ theorem, if X and Y are variables of the same dynamical system, the reconstructed attractors MX and MY are diffeomorphic to each other, as well as to the true attractor. In fact, if the two series represent a dimension in the state space, the knowledge of the shadow manifold MX , obtained from the variable X, can be used to estimate variable Y. Hence, we observed the similarity between the time series Yˆ , estimated from variable X, and the original time series Y, to quantify the imprint of variable X on variable Y. For each time series, we calculated the appropriate time delay τ as the minimum of the mutual information function, to ensure statistical independence between the values, in both linear and nonlinear terms. Moreover, we estimated the embedding dimension m, using the false nearest neighbour (FNN) method, as in [27]. Using time-delayed
Investigating the Physiology Behind Nose Thermal Response
81
embedding, from each variable X and Y we constructed the shadow manifolds, MX and MY . Afterwards, we identified the m+1 nearest-neighbours in MX over time. The number of neighbours depends on the maximum value of embedding dimension m between the two time series. The time index of these m+1 points in MX was used to obtain the corresponding points in MY . These points in MY were used to map a region of m+1 points around Yt . Finally, the estimate of each point of Y starting from the reconstructed manifold of X is found as follows:
Y t |MX =
m+1
wi Yti
(1)
i=1
where ωi are the weights obtained as the Euclidean distances between Xt and the nearestneighbours points (|| · || indicates the Euclidean distance in Rm ): ωi = ui ui wi = m+1
j=1
uj
Xt − Xti ui = exp − Xt − Xt1
(2) (3)
Finally, we computed the Pearson correlation coefficient between the original thermal signals and the corresponding estimates obtained using cross-mapping, and we will refer to it as ρ YY ˆ . Similarly, we computed the Pearson correlation coefficients between the original thermal signals and the HRV time series, ρ XY . 2.4 Statistical Analysis We performed an exploratory statistical analysis to evaluate the effect of the stress stimulus on the psychometric scores and the IR and cardiovascular dynamics. We used the non-parametric Wilcoxon sign-rank test to compare the STAY-Y1 and PS scores before and after the stressor. Analogously, we used the same test to compare each IR and HRV feature between Rest and Stress (α = 0.05). We controlled for the false discovery rate that results from multiple testing through the Benjamini, Krieger, and Yekutieli (FDR-BKY) correction [28]. In addition, we assessed the significance level of each correlation coefficient ρ YY ˆ using a non-parametric permutation test with 1000 repetitions (α = 0.05). The original thermal signal was divided into blocks of length equal to τ, which were randomly shuffled at each iteration to generate the null distribution. In this way, maximizing independence among blocks, we mitigate the effect of autocorrelation in the time series in causing spurious correlations between variables. Likewise, we assessed the significance levels of the Pearson correlation coefficients ρ XY using the same permutation test. Finally, a non-parametric Wilcoxon sign-rank test was used to test for possible differences between the correlation coefficient ρ YY ˆ related to the Rest and Stress sessions. Of note, the use of non-parametric statistical tests was justified by the non-gaussian distributions of samples.
82
F. Gioia et al.
3 Results The statistical analysis on both psychometric tests indicated that subjects re ported to be more stressed after the stimulation. In particular, the Wilcoxon sign-rank test on the PS levels reported a p-value equal to 0.01, and the p-value for the STAY-Y1 total scores was equal to 0.03. Regarding the physiological features, the mean temperature of the nose de creased significantly during Stress compared to Rest. Conversely, the mean of HRV increased significantly during Stress. The remaining HRV features, except the LF/HF ratio, decreased significantly compared to Rest, as shown in Fig. 2.
Fig. 2. Statistical comparison for each of the IR and HRV features extracted. Significant differences between Rest = R and Stress = S after the Wilcoxon signed rank test with FDR-BKY correction are highlighted with an asterisk (* = p < 0.05; ** = p < 0.01; *** = p < 0.001).
The results of the correlation between the original thermal signals and the HRV series ρ XY , and the correlation between the original thermal signals and their reconstructed version from the HRV time series ρ YY ˆ , are shown in Table 2. In particular, the median of the correlation coefficients ρ XY was 0.005 during Rest, and it was statistically significant in 17.86% of the subjects. During Stress, the median of the correlation coefficients ρ XY was equal to −0.123, and it was statistically significant in 60.71% of the subjects. The median of the ρ YY ˆ was equal to 0.7 during Rest and it increased to 0.78 during Stress. Note that, the associated p-values were significant for all the subjects (100%). Table 2. Non-parametric permutation test on correlation coefficient values ρ XY and ρ Y Y ˆ . Rest
Stress
ρ median
p-val < 0.05 (%)
ρ median p-
p-val < 0.05 (%)
ρ XY
0.005
17.86
−0.123
60.71
ρ YY ˆ
0.7
100
0.78
100
The boxplots in Fig. 3 show the values of the correlation coefficient ρ YY ˆ obtained after the application of the cross-mapping approach to the thermal signals of the nose
Investigating the Physiology Behind Nose Thermal Response
83
and the HRV time series. After the Wilcoxon signed rank test, the ρ YY ˆ values were significantly higher during Stress than Rest.
Fig. 3. Statistical comparison of ρ YY ˆ coefficient values obtained after the application of crossmapping method on thermal and HRV data, between Rest = R and Stress = S. Significant differences after the Wilcoxon signed rank test are highlighted with an asterisk (* = p < 0.05).
4 Discussions and Conclusions An understanding of the physiological processes driving changes in the facial temperature pattern could improve our thermography-based inferences in the field of psychophysiology. Greater emotional specificity would broaden the scope of IRT as a convenient non-contact method for monitoring changes in mental state. Although there are already hypotheses on the mechanisms that influence skin temperature, the quantification of these processes and their interplay is still lacking. This study reports an attempt to investigate the linear and non-linear relation between the thermal behaviour of the nose and cardiovascular activity, under both parasympathetic and sympathetic dominance. We used the cross mapping approach [20, 21], a nonlinear methodology based on chaos theory to quantify the agreement between HRV and nose thermal time series. Prior to the cross-mapping analysis, we tested the difference in the psycho metric scores reported before and after the stimulation. Similarly, we performed a comparison of thermal and HRV features between the two experimental sessions. The results confirmed that the protocol elicited the expected response, as the reported stress level increased and the physiological features varied significantly in agreement with the literature. Notably, nose temperature decrease during stress has been previously associated with sympathetic vasoconstriction of the subcutaneous vessels. The RMSSD and the PNN50 are temporal indexes of the vagal activity, as the HF power in the frequency domain. Their decrease is likely to reflect a decrease in the PNS activity during stress. Nonlinear analysis of the IR and HRV time series revealed a strong relation between the thermal and cardiovascular dynamics that was not obvious in the linear domain. Indeed, the linear correlation between the original HRV and thermal series was significant in the majority of the subjects during Stress. However, the correlation values ρ XY reported a median value as low as 0.005. On the other hand, the correlation values between the thermal signals of the nose and their estimated version obtained using the cross-mapping approach with the HRV time series were significant for all the subjects during both Rest and Stress, with a median value higher than 0.7. This nonlinear correlation values quantify
84
F. Gioia et al.
how the knowledge of the shadow manifold obtained from the time series of the HRV time series can be used to estimate values of the IR time series. Hence, we used the correlation value ρ YY ˆ as an index of the ‘imprint’ of the HRV time series on the thermal signal of the nose. Furthermore, the cross-mapping correlation values ρ YY ˆ were significantly higher during Stress compared to Rest. This outcome suggests that cardiovascular activity and nose temperature coupling are higher during elevated arousal conditions. Hence, our approach seems to corroborate the hypothesis that the stress-induced decrease in the nose temperature is related to sympathetic-induced vasomotor activity [12–14]. In our future studies, we will analyze the linear and nonlinear correlations be tween thermal signals from different regions of the face and several physiological signals related to the phenomena which are likely to underlie skin temperature modulation (i.e. heart rate variability, electrodermal and respiratory activity).
References 1. Ioannou, S., Gallese, V., Merla, A.: Thermal infrared imaging in psychophysiology: potentialities and limits. Psychophysiology 51, 951–963 (2014) 2. Cacioppo, J.T., Berntson, G.G., Klein, D.J., Poehlmann, K.M.: Psychophysiology of emotion across the life span. Annu. Rev. Gerontol. Geriatr. 17, 27–74 (1997) 3. Ekman, P., et al.: Universals and cultural differences in the judgments of facial expressions of emotion. J. Pers. Soc. Psychol. 53, 712 (1987) 4. Jerritta, S., Murugappan, M., Nagarajan, R., Wan, K.: Physiological signals based human emotion recognition: a review. In: 2011 IEEE 7th International Colloquium on Signal Processing and Its Applications, pp. 410–415 (2011) 5. Puri, C., Olson, L., Pavlidis, I., Levine, J., Starren, J.: StressCam: non contact measurement of users’ emotional states through thermal imaging. In: CHI’05 Extended Abstracts on Human Factors in Computing Systems, pp. 1725–1728 (2005) 6. Cardone, D., et al.: Driver stress state evaluation by means of thermal imaging: a supervised machine learning approach based on ECG signal. Appl. Sci. 10, 5673 (2020) 7. Perpetuini, D., et al.: Can functional infrared thermal imaging estimate mental workload in drivers as evaluated by sample entropy of the fNIRS signal? In: 8th European Medical and Biological Engineering Conference: Proceedings of the EMBEC 2020, November 29– December 3, 2020 Portorož, Slovenia, pp. 223–232 (2021) 8. Nˇemcová, A., et al.: Multimodal features for detection of driver stress and fatigue: review. IEEE Trans. Intell. Transp. Syst. 22, 3214–3233 (2021) 9. Akbar, F., et al.: Email makes you sweat: examining email interruptions and stress using thermal imaging. In: Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pp. 1–14 (2019) 10. Cho, Y., Bianchi-Berthouze, N., Oliveira, M., Holloway, C., Julier, S.: Nose heat: exploring stress-induced nasal thermal variability through mobile ther mal imaging. In: 2019 8th International Conference on Affective Computing and Intelligent Interaction (ACII), pp. 566–572 (2019) 11. Filippini, C., Spadolini, E., Cardone, D., Merla, A.: Thermal imaging based affective computing for educational robot. Multidiscip. Digit. Publ. Inst. Proc. 27, 27 (2019) 12. Cho, Y., Julier, S.J., Bianchi-Berthouze, N.: Instant stress: detection of perceived mental stress through smartphone photoplethysmography and thermal imaging. JMIR Mental Health 6, e10140 (2019)
Investigating the Physiology Behind Nose Thermal Response
85
13. Panasiti, M.S., et al.: Thermal signatures of voluntary deception in ecological conditions. Sci. Rep. 6, 1–10 (2016) 14. Or, C.K., Duffy, V.G.: Development of a facial skin temperature-based methodology for non-intrusive mental workload measurement. Occup. Ergon. 7, 83–94 (2007) 15. Shastri, D., Papadakis, M., Tsiamyrtzis, P., Bass, B., Pavlidis, I.: Perinasal imaging of physiological stress and its affective potential. IEEE Trans. Affect. Comput. 3, 366–378 (2012) 16. Goldberger, A.L., West, B.J.: Fractals in physiology and medicine. Yale J. Biol. Med. 60, 421 (1987) 17. Vuksanovi´c, V., Sheppard, L.W., Stefanovska, A.: Nonlinear relationship between level of blood flow and skin temperature for different dynamics of temperature change. Biophys. J. 94, L78–L80 (2008) 18. Kitney, R.: An analysis of the nonlinear behaviour of the human thermal vasomotor control system. J. Theor. Biol. 52, 231–248 (1975) 19. Di Credico, A., et al.: Estimation of heart rate variability parameters by machine learning approaches applied to facial infrared thermal imaging. Front. Cardiovasc. Med. 9 (2022) 20. Sugihara, G., et al.: Detecting causality in complex ecosystems. Science 338, 496–500 (2012) 21. Nardelli, M., Vanello, N., Galperti, G., Greco, A., Scilingo, E.P.: Assessing the quality of heart rate variability estimated from wrist and finger ppg: a novel approach based on cross-mapping method. Sensors 20, 3156 (2020) 22. Nardelli, M., Greco, A., Vanello, N., Scilingo, E.P.: Reliability of pulse rate variability in elderly men and women: an application of cross mapping approach. In: 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pp. 492–495 (2021) 23. Stroop, J.R.: Studies of interference in serial verbal reactions. J. Exp. Psychol. 18, 643 (1935) 24. Nirkin, Y., Masi, I., Tuan, A.T., Hassner, T., Medioni, G.: On face segmentation, face swapping, and face perception. In: 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018), pp. 98–105 (2018) 25. Gioia, F., Pascali, M.A., Greco, A., Colantonio, S., Scilingo, E.P.: Discriminating stress from cognitive load using contactless thermal imaging devices. In: 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pp. 608–611 (2021) 26. Tarvainen, M.P., Niskanen, J.-P., Lipponen, J.A., Ranta-Aho, P.O., Karjalainen, P.A.: Kubios HRV–heart rate variability analysis software. Comput. Methods Programs Biomed. 113, 210– 220 (2014) 27. Kennel, M.B., Brown, R., Abarbanel, H.D.: Determining embedding dimension for phasespace reconstruction using a geometrical construction. Phys. Rev. A 45, 3403 (1992) 28. Benjamini, Y., Krieger, A.M., Yekutieli, D.: Adaptive linear step-up procedures that control the false discovery rate. Biometrika 93, 491–507 (2006)
Comparing Valence-Arousal and Positive-Negative Affect Models of Affect: A Nonlinear Analysis of Continuously Annotated Emotion Ratings Andrea Gargano1,2(B) , Enzo Pasquale Scilingo1,2 , and Mimma Nardelli1,2 1 Dipartimento Di Ingegneria Dell’Informazione, Universitá Di Pisa, 56122 Pisa, Italy
[email protected] 2 Research Center “E. Piaggio”, 56122 Pisa, Italy
Abstract. The Affective Computing community commonly uses dimensional models of emotions to rate conscious emotional perceptions in emotion elicitation tasks. Although several structures of affect have been introduced in the literature, the valence and arousal dimensions have had the most impact. In this study, we compared Watson and Tellegen’s Positive-Negative Affect model to Russell’s Valence-Arousal plane. We used the publicly available Continuously Annotated Signals of Emotions (CASE) dataset, which provides ratings along the valence and arousal dimensions continuously annotated while watching video clips eliciting four emotions: scariness, amusement, relaxation, and boredom. We derived the Positive and Negative Affect time series from the valence and arousal time series through a 45° rotation of Russell’s plane. We calculated the median values and Fuzzy Entropy for each time series and video clip to investigate their linear and nonlinear dynamics. Our analysis showed that Watson and Tellegen’s model had fewer statistically significant differences between emotions than Russell’s model when considering the median values. However, when investigating the dynamic evolution of perceptions, the Positive Affect dimension showed the highest discriminative power, identifying the time series traced during the boring stimuli as the most regular and statistically different from all others. Our findings suggest that further acquisitions of continuously annotated ratings in several experimental settings, and the investigation of the nonlinear coupling between more dimensions, could significantly improve real-time emotion recognition. Keywords: Affective computing · Continuously annotated ratings · Emotion discrimination · Valence-arousal plane · Positive-negative affect model · Fuzzy entropy · Nonlinear time series analysis
1 Introduction In the Affective Computing field, emotion elicitation and recognition are topics strictly influenced by the complexity of the definition of consciously experienced affective states [24]. According to the previous literature, models of emotions were broadly organised © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 86–95, 2024. https://doi.org/10.1007/978-3-031-49062-0_10
Comparing Valence-Arousal and Positive-Negative
87
into two categories [12, 13, 23, 24]: discrete and dimensional models. The former essentially aimed at determining a set of primary and innate emotions, leading to prototypical responses when elicited [9, 12]. On the other hand, dimensional models emerged from exploring core dimensions [24, 31] that described more than a single discrete emotional state at once. Furthermore, more recent and sophisticated models were proposed, e.g. the “appraisal-based models”, but they are still unsuited for the experimental assessment of emotions [13]. Currently, the most widespread dimensional model is the Circumplex model of Affects proposed by Russell [23]. It consists of a plane with two orthogonal and bipolar dimensions: valence and arousal. The former denotes the hedonic dimension, i.e. the degree of the pleasantness of the perceived emotion, ranging from unpleasant to pleasant. The latter describes the intensity of the emotion felt that is represented by the arousal degree, ranging from low to high arousal. Alternative interpretations of the structure of affects were proposed [5], such as the Positive-Negative Affect model defined by Watson and Tellegen [31]. They proposed two novel dimensions, namely Positive Affect (thereinafter PA) and Negative Affect (thereafter NA), which they considered to be the subjective expression of the two fundamental behavioural systems of approach and with drawal, respectively [31]. Therefore, paralleling Russell’s work, they portrayed a novel Affect Circumplex entailing four bipolar dimensions spaced 45° apart [31]: pleasantness, PA, engagement, and NA. Being largely independent and negatively correlated, the authors emphasised the PA and NA as essential dimensions for the structure of affects, which they measured throughout the PANAS scale, one of the most widely used psychometric scales by the Affective Computing community [30]. Throughout this work, we will refer to the circumplex model defined by PA and NA as the PA-NA plane. Lately, innovative annotation tools deployed in experimental protocols have allowed participants to rate perceived emotions in more than a single dimension at once. In particular, the experimental setup of the Continuously Annotated Signal of Emotion (CASE) dataset [26] was specifically devised to continuously record self-assessed ratings of arousal and valence dimensions in real-time, during video elicitation sessions. In a previous study [11], we investigated the dynamics of the arousal annotation data considering two video stimuli only, the relaxing and the scary. An experimental comparison between the time-continuous emotion ratings in the valence-arousal plane and the rearranged values in the PA-NA plane has not yet been presented in the literature. In this study, we used nonlinear time series analysis techniques to investigate the dynamics of the CASE dataset time-varying ratings in the valence-arousal plane and their transformation in the PA-NA plane. Previous works in the domain of physiological time series analysis highlighted the importance of taking advantage of nonlinear analysis approaches to account for the dynamics of these signals and characterise the autonomic [17–19] and the central nervous system activities for emotion recognition tasks [10]. We investigated the dynamics of all four dimensions of emotions (i.e., valence, arousal, PA, and NA) using Fuzzy Entropy (hereinafter FuzzyEn), which was developed to measure physiological time series regularity [7]. Moreover, we calculated the median of each time series as a reference averaged time-domain metric. Then, we focused on the statistical differences between the four emotions (scariness, amusement, relaxation, and boredom) elicited by the CASE video stimuli.
88
A. Gargano et al.
2 Materials and Methods 2.1 The Continuously Annotated Signal of Emotion (CASE) Dataset The publicly-available CASE dataset [25, 26] provides continuously annotated ratings of emotions, recorded from 30 young adults (15 females, aged 25.71 ± 3.1 years, and 15 males, aged 28.6 ± 4.8 years). The arousal and valence ratings were collected by asking each participant to view eight emotional video clips with duration in the range [119, 197] s and continuously rate his/her emotional state in real-time [25, 26]. The videos were singled out to induce four distinct emotional states (relaxation, scariness, amusement, and boredom, with two clips for each emotion) and presented in pseudo-randomized order. The CASE dataset is distinct from other datasets furnishing continuous ratings of emotions [1, 21, 22] since each subject evaluated its emotional state along the valence and arousal dimensions in real-time while viewing the emotion eliciting video clips. A custom joystick-based interface [3, 27] allowed a simultaneous annotation along both affective dimensions. Specifically, the joystick interface was connected to a graphic interface that appeared on the upper right corner of the screen where the video clip was played [3]. The graphic interface was based on Russell’s plane [23] and was enhanced by placing the icons of the Self-Assessment Manikin [6] on each axis to simplify the annotation task. Before the elicitation, a practice session of five short videos endowed participants with familiarising themselves with the task and the annotation interface. The annotation data were acquired with a sampling frequency of 20 Hz [26]. 2.2 The Positive and Negative Affect (PA-NA) Plane From the original valence and arousal data, we generated a novel dataset made by the PA and NA annotation data, to compare the four annotated data types (arousal, valence, PA, and NA) and investigate their emotion discrimination capability. Specifically, in [31] a novel structure of affect was proposed based on two dimensions: PA and NA. A unique affect circumplex, containing four bipolar axes (i.e., pleasantness, engagement, PA, and NA), was depicted with dimensions spaced 45° apart [31]. From a geometric perspective, since the PA-NA plane concerns a 45° rotation of the valence-arousal plane, our idea was to apply a rotation to the original data collected in the valence and arousal plane to make them fit the PA-NA plane. Valence and arousal data were rated in the [0.5, 9.5] range. Therefore, to obtain the PA and NA data, we first removed the origin (i.e., [6, 6]) of the valence-arousal plane. To rotate the data, we applied a rigid transformation to each data vector of valence and arousal time series through a standard rotation matrix, with rotation angle θ equal to 45°, according to: pa(t) cos θ sin θ v(t)− v0 (1) = a(t)− a0 na(t) − sin θ cos θ with pa(t) and na(t) being the time series along the PA and NA dimensions, v(t) and a(t) represent the original valence and arousal time series, and v0 and a0 are the values of the origin for the valence and arousal time series, respectively. We applied this operation
Comparing Valence-Arousal and Positive-Negative
89
for all eight video types by coupling the annotated valence and arousal time series. Therefore, for each video type, through this transformation, we gained two additional annotation types along the PA and NA dimensions. 2.3 Phase Space Reconstruction and Fuzzy Entropy Analysis The first step for computing nonlinear indexes of the annotated ratings was the phase space reconstruction for each of the four dimensions. We used the Takens embedding theorem [28] to reconstruct the so-called embedded vectors, which describe each time series’s trajectory in the phase space. We computed two parameters for each time series to construct the relative embedded vectors: the time delay τ and the embedding dimensions m. The first parameter is the time lag to plot the time series against itself; the second one represents the dimension of the phase space, i.e., the dimension of the embedded vectors. According to [2], we calculated τ as the first minimum of the auto-mutual information function, estimated through the kernel density estimation approach described in [29]. Additionally, we computed the value of m applying the False Nearest Neighbors (FNN) algorithm, proposed in [14]. Based on the parameters τ and m, we defined the states of the valence, arousal, PA, and NA dynamical systems in their own phase spaces throughout the coordinates of their embedded vectors. Specifically, from the original N dimensional time series x = [x(1), x(2),…, x(N)], we computed the N − (m − 1)τ embedded vectors in Rm . The i-th embedded vector ui was calculated as ui = [x(i), x(i + τ ),…, x(i + (m − 1)τ )], with i ∈ [1, N − (m − 1)τ ]. Following the reconstruction of the phase spaces of the four affective dimensions, we characterised their information content by employing information theoretic approaches. We used the FuzzyEn as a measure of regularity of the reconstructed dynamics [7]. In contrast to other entropy indexes, such as Sample Entropy and Approximate Entropy [20], FuzzyEn is not based on a binary comparison with a fixed threshold to estimate the similarity between points in the phase space. It is rather established on the broader concept of fuzzy sets to measure the closeness between points. Consequently, each distance between two vectors contributes to the estimate of trajectory similarity. Practically, the first step for the computation of the FuzzyEn relies on the computation of the Chebyshev distance between each pair of embedded vectors ui and uj in the phase space, with i = j to exclude self-matches, as follows:
(2) d ui , uj = max xi+(k−1)τ − xj+(k−1)τ k=1,...,m
An exponential function is employed as a membership degree function, which assigns a value in the continuous range [0, 1] for each distance value d(ui , uj ). Specifically, this distance value is used to compute the similarity degree Dm ij (n, r) between embedded vector ui to uj , according to the following equation: with m being the embedding dimension; n and r are parameters linked to the width and the gradient of the boundary of the exponential function, respectively. According to previous n (3) Dijm (n, r) = exp(− d m (ui , uj /r)
90
A. Gargano et al.
work in physiological time series analysis [7], for our preliminary analysis we set r equal to 20% of the standard deviation of each time series and n equal to 2. Based on the FuzzyEn algorithm, the sample correlation measure Am (n, r) is defined as the similarity degree value accounting for all vectors in the phase space, normalized by the total number of vectors N − mτ, as in the following equation: ⎡ ⎤ N −mτ N −mτ 1 1 ⎣ Am (n, r, τ ) = Dijm (n, r)⎦ (4) N − mτ N − mτ − 1 i=1
i=1,i=j
Afterwards, the value of the embedding dimension is increased from m to m + 1 and the values of Dijm+1 (n, r) are computed for each new pair of (m + 1)-dimensional vectors in the phase space, according to Eq. (3). Then, as shown in Eq. (4), we computed the new value of the sample correlation measure Am+1 (n, r). Finally, we computed the value of FuzzyEn according to the following formula: FuzzyEn(n, r, m, τ ) = −ln(
Am+1 (n, r) ) Am (n, r)
(5)
As a reference measure in the time domain, to compare the emotion discrimination capability along each dimension, we computed the median of the time series as an average measure. 2.4 Statistical Analysis In this work, for each of the four annotation data types (i.e., arousal, valence, PA, NA), we averaged across the same subject the FuzzyEn and median values computed for the two videos with the same emotional content. Therefore, we obtained a single subjectdependent measurement of FuzzyEn and median for each emotion type. For all four annotation data types, we tested independently the FuzzyEn and the median through a within-subject statistical comparison between the four emotions. We performed a non-parametric Friedman test to check any difference between the medians of the four different emotion types for each measure and each annotation data type. If we found a significant difference, we applied the Wilcoxon signed-rank test for paired samples as the multiple comparison test. We set the statistical significance level at α = .05 and applied the Bonferroni correction when testing for multiple comparisons. We used non-parametric statistical tests due to the non-gaussianity of the sample distributions, demonstrated by testing each sample with the Shapiro-Wilk test. According to the signal quality, all the statistical tests were performed on 26 subjects. We executed all the analyses with the software Matlab (Release 2021b, Mathworks Inc., Natick, MA).
3 Results The FuzzyEn presented the lowest median (± median absolute deviation, i.e., MAD) value for the boring stimulation compared to all the other emotion stimulation types, as highlighted by the violin plots in Fig. 1a. In particular, for the arousal annotation data,
Comparing Valence-Arousal and Positive-Negative
91
the FuzzyEn of the boring stimulation (0.181 ±0 .085) resulted significantly lower than the amusing (0.302 ± 0.090, p = 0.0050) and the scary (0.263 ± 0.041, p = 0.0476) ones. Similarly, for the valence annotation data, the FuzzyEn for the boring (0.193 ±0 .062) stimulation was significantly lower than the amusing (0.328 ±0 .065, p =0 .0096) and the scary (0.254 ±0 .007, p = 0.0045) clips. Regarding the PA data, the FuzzyEn for the boring (0.138 ±0 .052) stimulation was significantly lower than all the other induced emotions: fear (0.243 ± 0.053, p = 0.0007), amusement (0.259 ± 0.053, p = 0.0002), and relaxation (0.229 ± 0.078, p = 0.0176). However, for the annotations along the NA dimension, the FuzzyEn for the boring (0.160 ± 0.058) stimulation came about to be significantly lower than the amusing (0.233 ±0 .029, p = 0.0131) only. Concerning the analysis of the median, by looking at Fig. 1b we can appreciate that for the arousal data, all pairwise comparisons were significantly different, whereas, for the valence data, the only non-significant comparison was between amusing (6.308 ± .539) and relaxing (5.879 ± 0.733) stimulations. However, a different picture was unveiled when dealing with the median of the derived annotation data PA and NA. More in detail, for the median of the PA annotation data, the amusing stimulation is significantly higher than all the others, as well as the boring has the lowest median (−0.673 ±0 .695) compared to the others: relaxing (0.056 ± 0.431, p = 0.0022), scary (0.391 ± 0.407, p = 0.0002), and amusing (1.140 ± 0.546, p .0001). Regarding the NA data, the fear-inducing stimulation presents a higher median (2.248 ± .870) than the others: the amusing (−0.719 ± 0.373, p 0.0001), the relaxing (−0.876 ±0 .513, p 0.0001), and the boring (−0.823 ±0 .401, p 0.0001), but there is no difference among the last three.
4 Discussion and Conclusions In this computational study, we conducted a preliminary investigation into the dynamics of four conscious emotions that were collected through continuous recordings of selfassessed ratings during an emotion-eliciting task. We utilized the arousal and valence annotated signals provided by the CASE dataset to derive the Positive Affect (PA) and Negative Affect (NA) dimensions due to their mutual connections [30, 31]. In our previous work, we explored the possibility of distinguishing between relaxation and scariness using entropy indexes [11]. However, no previous study has compared emotions based on a nonlinear analysis of their dynamics according to two different emotion models: Russell’s valence arousal plane [23] and Watson and Tellegen’s PA-NA plane [31]. To assess the dynamics’ regularity, we utilized FuzzyEn, given its robustness against short time series. Pairwise comparisons of the PA and NA time series highlighted the possibility of successfully distinguishing boredom from the other three emotions (scariness, amusement, and relaxation) using the PA data. When comparing the statistical findings obtained by applying nonlinear analysis techniques to the four emotion dimensions, PA emerged as the unique scale capable of completely discriminating one emotion (boredom). In contrast, the FuzzyEn for the valence and arousal dimensions did not differentiate between relaxation and boredom. We also analyzed the median of the annotated time series. Although almost all four emotions were differentiated by the valence and arousal annotation median, the same did not hold for the PA and NA annotations. Specifically,
92
A. Gargano et al.
Fig. 1. Violin plots depicting the dispersion of Fuzzy Entropy (FuzzyEn) a and median b values for the arousal (top left), valence (top right), positive arousal (PA, bottom left), and negative arousal (NA, bottom right) annotation data. Values reported were obtained by averaging for the same subject two entropy indexes a and median values b, calculated by starting from the annotated time series (i.e., arousal, valence, PA, and NA) of each of the two videos inducing the same emotion category. Statistically significant p-values, corrected by the number of multiple comparisons, are reported in accordance with the following legend: ∗ p ≤ .050; ∗∗ p ≤ .010; ∗∗∗ p ≤ .001.
Comparing Valence-Arousal and Positive-Negative
93
the median of PA annotations demonstrated superior discrimination power compared to NA, with the latter being useful only for differentiating scariness. Therefore, combining the average and the nonlinear entropy index would increase the emotion discrimination capability in the PA-NA plane. One limitation of this study is the data transformation from the valence arousal plane to the PA-NA plane. Specifically, users were trained to rate their emotional state in the valence-arousal plane only. Since the CASE dataset did not contain annotations along the PA and NA dimensions, we derived these data. Additionally, the original annotated data were collected in a rectangular valence-arousal plane, constraining the boundaries of the rotated space. We plan to validate our results in future studies with other validated entropy metrics (e.g., Distribution Entropy [16]). This will allow us to thoroughly characterize the dynamics of these novel annotation signals and unleash their potential for real-time emotion classification tasks. An essential step in that direction would be to investigate the emotion discrimination capabilities by coupling annotated signals along different dimensions in a multivariate analysis. Specifically, several psychological models of emotion suggest that bidimensional models might fail to capture subtle differences between emotions. Therefore, as in [8], we could couple more than two dimensions. Furthermore, according to [31], we applied a rotation to the valence-arousal plane of 45 degrees. However, our future research will investigate how different rotation angles impact the emotion discrimination capabilities of these annotated data. Given the considerable amount of studies showing the discriminative power of nonlinear analysis of physiological signals [7, 10, 15–19], we believe that combining this information with the study of annotated signals could remarkably disclose still unnoticed connections between consciously experienced emotions and unconscious physiological processes. Acknowledgement. The research leading to these results has received partial funding from European Union Horizon 2020 Programme under grant agreement n° 824153 of the project “POTIONPromoting Social Interaction through Emotional Body Odours” and by the Italian Ministry of Education and Research (MIUR) in the framework of the FoReLab project and CrossLab project (Departments of Excellence).
References 1. Abadi, M.K., Ramanathan, S., Kia, S.M., Avesani, P., Patras, I., Sebe, N.: De caf: Meg-based multimodal database for decoding affective physiological responses. IEEE Trans. Affect. Comput. 6, 209–222 (2015) 2. Abarbanel, H.D., Brown, R., Sidorowich, J.J., Tsimring, L.S.: The analysis of observed chaotic data in physical systems. Rev. Mod. Phys. 65(4), 1331 (1993) 3. Antony, J., Sharma, K., Castellini, C., van den Broek, E.L., Borst, C.W.: Continuous affect state annotation using a joystick-based user interface. In: Proceedings of Measuring Behavior 2014: 9th International Conference on Methods and Techniques in Behavioral Research, pp. 500–505 (2014) 4. Barrett, L.F., Russell, J.A.: The structure of current affect: controversies and emerging consensus. Curr. Dir. Psychol. Sci. 8(1), 10–14 (1999)
94
A. Gargano et al.
5. Bradley, M.M., Lang, P.J.: Measuring emotion: the self-assessment manikin and the semantic differential. J. Behav. Ther. Exp. Psychiatry 25(1), 49–59 (1994) 6. Chen, W., Wang, Z., Xie, H., Yu, W.: Characterization of surface emg signal based on fuzzy entropy. IEEE Trans. Neural Syst. Rehabil. Eng. 15(2), 266–272 (2007) 7. Eerola, T., Vuoskoski, J.K.: A comparison of the discrete and dimensional models of emotion in music. Psychol. Music 39(1), 18–49 (2011) 8. Ekman, P.: An argument for basic emotions. Cogn. Emot. 6(3–4), 169–200 (1992) 9. García-Martínez, B., Martinez-Rodrigo, A., Alcaraz, R., Fernández-Caballero, A.: A review on nonlinear methods using electroencephalographic recordings for emotion recognition. IEEE Trans. Affect. Comput. 12(3), 801–820 (2019) 10. Gargano, A., Scilingo, E.P., Nardelli, M.: The dynamics of emotions: a preliminary study on continuously annotated arousal signals. In: 2022 IEEE International Symposium on Medical Measurements and Applications (MeMeA), pp. 1–6. IEEE (2022) 11. Grandjean, D., Sander, D., Scherer, K.R.: Conscious emotional experience emerges as a function of multilevel, appraisal-driven response synchronization. Conscious. Cogn. 17(2), 484–495 (2008) 12. Gunes, H., Schuller, B., Pantic, M., Cowie, R.: Emotion representation, analysis and synthesis in continuous space: a survey. In: 2011 IEEE International Conference on Automatic Face & Gesture Recognition (FG), pp. 827–834. IEEE (2011) 13. Kennel, M.B., Brown, R., Abarbanel, H.D.: Determining embedding dimension for phasespace reconstruction using a geometrical construction. Phys. Rev. A, Atom. Mol. Opt. Phys. 45(6), 3403–3411 (1992) 14. Lake, D.E., Richman, J.S., Griffin, M.P., Moorman, J.R.: Sample entropy analysis of neonatal heart rate variability. Am. J. Physiol.-Regul. Integr. Comp. Physiol. 283(3), R789–R797 (2002) 15. Li, P., Liu, C., Li, K., Zheng, D., Liu, C., Hou, Y.: Assessing the complexity of short-term heartbeat interval series by distribution entropy. Med. Biol. Eng. Compu. 53(1), 77–87 (2015) 16. Nardelli, M., Lanata, A., Bertschy, G., Scilingo, E.P., Valenza, G.: Heartbeat complexity modulation in bipolar disorder during daytime and nighttime. Sci. Rep. 7(1), 1–11 (2017) 17. Nardelli, M., Valenza, G., Greco, A., Lanata, A., Scilingo, E.P.: Recognizing emotions induced by affective sounds through heart rate variability. IEEE Trans. Affect. Comput. 6(4), 385–394 (2015) 18. Nardelli, M., Valenza, G., Greco, A., Lanata, A., Scilingo, E.P., Bailón, R.: Quantifying the lagged poincaré plot geometry of ultrashort heart rate variability series: automatic recognition of odor hedonic tone. Med. Biol. Eng. & Comput. 58(5), 1099–1112 (2020) 19. Richman, J.S., Moorman, J.R.: Physiological time-series analysis using approximate entropy and sample entropy. Am. J. Physiol.-Hear. Circ. Physiol. 278(6), H2039–H2049 (2000) 20. Ringeval, F., Schuller, B., Valstar, M.F., Gratch, J., Cowie, R., Scherer, S., Mozgai, S., Cummins, N., Schmitt, M., Pantic, M.: Avec 2017: real-life depression, and affect recognition workshop and challenge. In: Proceedings of the 7th Annual Workshop on Audio/Visual Emotion Challenge, pp. 3–9 (2017) 21. Ringeval, F., Sonderegger, A., Sauer, J.S., Lalanne, D.: Introducing the recola multimodal corpus of remote collaborative and affective interactions. In: 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pp. 1–8 (2013) 22. Russell, J.A.: A circumplex model of affect. J. Pers. Soc. Psychol. 39(6), 1161–1178 (1980) 23. Russell, J.A.: Core affect and the psychological construction of emotion. Psychol. Rev. 110(1), 145 (2003) 24. Sharma, K., Castellini, C., Stulp, F., Van den Broek, E.L.: Continuous, real-time emotion annotation: a novel joystick-based analysis framework. IEEE Trans. Affect. Comput. 11(1), 78–84 (2017)
Comparing Valence-Arousal and Positive-Negative
95
25. Sharma, K., Castellini, C., van den Broek, E.L., Albu-Schaeffer, A.O., Schwenker, F.: A dataset of continuous affect annotations and physiological signals for emotion analysis. Sci. Data 6 (2019) 26. Sharma, K., Wagner, M., Castellini, C., van den Broek, E.L., Stulp, F., Schwenker, F.: A functional data analysis approach for continuous 2-d emotion annotations. Web Intell. 17, 41–52 (2019) 27. Takens, F.: Detecting strange attractors in turbulence. In: Rand, D., Young, L.S. (eds.) Dynamical Systems and Turbulence, Warwick 1980, pp. 366–381. Springer, Berlin, Heidelberg (1981). https://doi.org/10.1007/BFb0091924 28. Thomas, R.D., Moses, N.C., Semple, E.A., Strang, A.J.: An efficient algorithm for the computation of average mutual information: validation and implementation in matlab. J. Math. Psychol. 61, 45–59 (2014) 29. Watson, D., Clark, L.A., Tellegen, A.: Development and validation of brief measures of positive and negative affect: the panas scales. J. Pers. Soc. Psychol. 54(6), 1063 (1988) 30. Watson, D., Tellegen, A.: Toward a consensual structure of mood. Psychol. Bull. 98(2), 219 (1985) 31. Watson, D., Wiese, D., Vaidya, J., Tellegen, A.: The two general activation systems of affect: Structural findings, evolutionary considerations, and psychobiological evidence. J. Pers. Soc. Psychol. 76(5), 820 (1999)
Postprandial Peak Identification from Continuous Glucose Monitoring Time Series Aikaterini Archavli1 , Harpal Randeva2,3 , and Natasha Khovanova1,2(B) 1 School of Engineering, University of Warwick, Library Rd, Coventry CV4 7AL, UK
[email protected]
2 University Hospital Coventry and Warwickshire, Clifford Bridge Rd, Coventry CV2 2DX, UK 3 Warwick Medical School, University of Warwick, Gibbet Hill Rd, Coventry CV4 7HL, UK
Abstract. Continuous glucose monitors (CGMs) have been mainly used in medical applications to monitor blood glucose and to control insulin doses in Type 1 diabetes (T1D) patients. CGMs are becoming popular in people without diabetes and with Type 2 diabetes (T2D) due to their rising commercial availability and effectiveness. They are a useful tool for understanding individuals’ dynamic blood responses to food. However, before such information can be extracted for further analysis, the peaks must be selected automatically. Published works have focused on detecting the onset of meal intakes and calculating their carbohydrate load to assist glucose control in T1D management. This work aims to develop a threshold-based algorithm for entire postprandial peak identification, including starting and endpoints, from data obtained from people with different glucose tolerance levels. The algorithm achieved promising performance using an individual threshold-based approach, with recall and precision rates of 0.84 and 0.85, respectively. Keywords: Diabetes Mellitus · Blood glucose · Peaks · Algorithms
1 Introduction Diabetes mellitus (DM) is a general term characterising a group of diseases that lead to prolonged hyperglycaemia, including Type 1 (T1D), Type 2 (T2D) and gestational diabetes mellitus. The most common type of diabetes is T2D when the body cannot effectively use the insulin it produces. According to the International Diabetes Federation (IDF), the global diabetes prevalence in adults aged between 20-79 years was estimated at 9.3%, rising to 10.2% by 2030 and 10.9% by 2045 [1]. However, in the 2022 edition of the IDF Atlas, the prevalence of diabetes in 2021 was 10.5% exceeding the previous projections. The latest predictions estimate the prevalence by 2045 to increase to 12.2%. The increasing prevalence emphasises why DM continues to be a significant global concern [2]. Continuous glucose monitoring (CGM) has led to a clinical shift in managing DM, with several studies showcasing that controlled behaviour specific to glucose monitoring © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 96–107, 2024. https://doi.org/10.1007/978-3-031-49062-0_11
Postprandial Peak Identification from Continuous
97
has been instrumental in managing glycemic levels [1, 3, 4]. CGM is used mainly for monitoring T1D; however, the use of CGM is expanding, with promising results in managing T2D, in patients with insulin treatment [5]. The benefits of applications of CGM in people developing DM (pre-DM), i.e. when their blood glucose levels are higher than normal but not high enough yet to be diagnosed as T2D, and in individuals without DM, are outlined in recent studies [6, 7]. CGM is considered a valuable signal tracker of glucose concentration. With the increasing use of CGM, effectively utilising all available information is essential for broad applications. One of the current topics of interest is detecting food intake in CGM signals. Meal detection is currently an area of interest for artificial pancreas (AP) devices (closed loop system) for T1D [8]. AP systems require users to provide information regarding their meal intakes as input or infer meals from CGM data so that the system can adjust the insulin delivery accordingly [9]. Various approaches for meal identification can be found in the literature utilising either simulation or real-life data. Several publications followed threshold-based approaches for meal identification and calculation of the carbohydrate load of the meal. Moreover, model-based techniques have also been utilised as well as a combination of threshold and model-based approaches [9–12]. Recently, meal identification in healthy individuals has been of interest [13] to validate self-reports of food intake timing. These approaches focussed mainly on identifying the times of meal intakes for no-DM individuals and/or calculating carbohydrate loads and the impact of physical activities to improve insulin treatment in a closed-loop system for T1D. However, no similar approaches were implemented for T2D or pre-DM cases or focussed on identifying the entire postprandial responses. An automatic or semi-automatic algorithm for extracting entire postprandial peaks can benefit large datasets when manual selection is not feasible. Such algorithms could be integrated with portable devices like the peak extraction algorithms for other signals, e.g. ECG. However, there are challenges in developing fully automated procedures, including the unavailability of public data, absence or inaccurate information on time intakes and composition of meals, unquantified effects of stress and physical activities, missing data, and inaccuracy of CGM devices. These data limitations restrict the development of fully automated artificial intelligence methods and require semi-automated algorithms involving human expert knowledge to help cope with complex data. This work aims to develop a threshold-based algorithm for entire postprandial peak identification from data obtained from people with different levels of glucose control and investigate its usefulness and robustness in relation to individual blood glucose variations.
2 Participants and Data Collection The experimental data consist of blood glucose profiles from eight subjects under two different living conditions. The subjects spend one day in the research site at the Human Metabolism Research Unit (HMRU) where they were confined to a small observation room (metabolic chamber) with controlled food and exercise routines (controlled environment). The rest of the days were spent under normal living conditions, following their normal daily routine while keeping a diary on food intake and exercise (free-living environment). The data were collected for 3 groups of participants: subjects without DM
98
A. Archavli et al.
(no-DM group), subjects diagnosed with T2D (DM group) and subjects with pre-DM (pre-DM group). The experimental data were collected in accordance with the Ethical approval from the National Health Authority Research Ethics Committee, North West—Liverpool East, United Kingdom (ID:17/NW/0277). The blood glucose profiles were recorded by the Medtronic iPro2 continuous glucose monitors (CGM) for seven days with a sampling period of five minutes. The first 24 h of data were excluded from the analysis due to inaccuracies created by the inflammatory response at the insertion site [14]. Besides glucose profiles, baseline data, such as age, gender, body mass index (BMI), glycaemic status (no-DM, pre-DM, T2D) and treatment regimen was also recorded. The treatment regimen is expected to be consistent and followed throughout the study. In total, 159 peaks were recorded, with 68, 41 and 50 peaks corresponding to the no-DM, pre-DM and T2D groups.
3 Algorithm Description 3.1 Algorithm Development To identify peaks in the glucose profiles, the proposed algorithm utilises a local maxima approach using thresholds for the height and timing of the peaks, adjusted to individual and group characteristics. The development of the algorithm was based on the peaks observed during the time the subjects spent in the controlled environment. The meals were specifically designed and followed by the participants while they were under controlled conditions. Each meal was consumed at a specific time during the day and was scaled accordingly to either body weight or calorie intake. For certain subjects, specific controlled meals had to be excluded due to subsequent meal intakes that were not controlled, and for one subject, multiple meals were considered as controlled as the subject followed the recommended meals under free-living conditions. In the next stage, the algorithm was tested on the glucose profiles of the same subjects in free-living conditions, where food timings and type of meals were not restricted or controlled but was recorded by the subjects in their food diaries [14]. In real-life data, multiple meals could occur very close to each other. For example, a subject can eat a snack or a dessert after finishing a meal. To account for this, if the data contained subsequent meals and they corresponded to the same glucose response, the meals were considered as one combined meal. As a result, a total of 30 glucose peaks were used for the development of the algorithm and 129 for the testing of the algorithm. 3.2 Algorithm Stages 1. Identification of the peaks corresponding to food intakes In this step, the peak’s starting point (Fig. 1a) was selected by searching and identifying the zero crossing point in the signal. For each starting point under consideration, the algorithm also searched for the following maximum point (Fig. 1a). Then the distance between the starting and maximum points also called the peak’s height, was considered at each step, and if it satisfied the specific condition below (item 2), the peak was selected.
Postprandial Peak Identification from Continuous
99
2. Peak height selection. The height for a peak to be selected must be greater than the specific thresholds for each group. The thresholds were selected by first calculating the heights of all peaks of the controlled meals per individual. After all the heights were calculated per individual (Fig. 2), the minimum values of the heights were extracted and then averaged across each group, i.e., one averaged height value for the no-DM group, one for the pre-DM group and one for the T2D group was selected. To avoid too high and too low values in the pre-DM and T2D group, the obtained average values were further averaged with the threshold of 1.1 mmol/L, which was used by Zhang et al. [15] for manual peak selection in the T2D group. As a result, the thresholds were set to 0.8 mmol/L for the no-DM group, 1.2 mmol/L for the pre-DM group and 1.3 mmol/L for the T2D group, and all the peaks higher than these thresholds were selected for further consideration. 3. Identification of the end point (Fig. 1a) of the glucose peak. The following conditions were used: – The duration of a peak is affected by meal composition, quantity and timing of the food intake, and it differs across individuals with and without DM [16]. Therefore, two timeframe thresholds were selected, one for the no-DM group (4 h) and one for the pre-DM and T2D groups (6 h). – The glucose meal response can include a double peak, i.e., a second elevation of glucose levels after the first peak. The second elevation occurs within a specific timeframe, usually between 90 and 120 minutes after a food intake [17] as seen in Fig. 1b. We assumed that if a ratio of the height of the second peak to the height of the first peak is higher than 0.33, such oscillation is unrealistic, and the second peak must be considered as a separate event, not a part of a double peak [15]. – For the double peaks, an additional threshold was used to consider if the second elevation was a separate event. This condition utilises the start time of the second peak in relation to the first peak. The threshold is adjusted accordingly for the no-DM, pre-DM and T2D groups. For the no-DM group, the timeframe of 40 minutes after the maximum point of the first peaks was selected; for pre-DM and T2D, the timeframe of 50 min was selected. The choice of threshold was adjusted accordingly with the timeframe of the second elevation from Cheng et al. [17] and the observations from the dataset. 4. Additional considerations. We assumed no food consumption had happened overnight, i.e. from midnight to 6 am.
3.3 Algorithm Assessment The algorithm was assessed using the following standard measures: – The number of the true positive (TP) peaks, i.e., the number of peaks detected by the algorithm that corresponded to the actual meal intakes – The number of false positive (FP) peaks, i.e., when detection was triggered, but no meal occurred – The number of false negative (FN) peaks, i.e., when a meal occurred, but no detection was triggered.
100
A. Archavli et al.
Fig. 1 Glucose profiles of a subject from no-DM group: (a) peak’s starting, maximum and end points, (b) double peak after a meal intake
Fig. 2 Boxplots of the identified peak heights per individual (S1, S2, S3, etc) in the controlled environment.
– True Positive Rate (TPR), or recall, representing the proportion of actual positive peaks that were identified correctly: Recall =
TP TP + FN
(1)
– Precision describing the proportion of identified peaks that were actually correct: Precision =
TP . TP + FP
(2)
4 Results and Discussion 4.1 Algorithm Performance Under Controlled Conditions (Dataset 1) Overall, 90% of the total number of controlled peaks were correctly identified by the algorithm leaving 10% of peaks undetected. In Fig. 3a and b the number of TP and FN peaks are presented per subject and group, respectively. The recall is calculated using Eq. (1) for each subject and group. The greatest recall value of 1 was obtained in the preDM group, where all the glucose peaks were identified correctly and no FP values were
Postprandial Peak Identification from Continuous
101
observed. This result was followed by a recall value of 0.93 for the no-DM group and 0.78 for the T2D group. In both no-DM and T2D groups, the recall rates were affected mainly by FN peaks identified in specific subjects, whereas five out of eight subjects had a perfect recall score of 1.
Fig. 3 Bar plots representing the algorithm results for each subject (a) and each group (b) for the controlled peaks. The blue bars represent the total number of peaks, the red bars correspond to the TP peaks and the yellow bars represent FN peaks.
4.2 Algorithm Performance Under Free-Living Conditions (Dataset 1) The overall recall rate of the algorithm for the peaks under free-living conditions was 0.78. The greatest recall rate was, again, observed in the pre-DM group, where it dropped from 1 (in the controlled conditions) to 0.88 in the free-living environment. The T2D group had the same recall (0.8) as in the controlled environment, whereas the recall for the no-DM group dropped from 0.93 to 0.75. Figure 4a and b present the number of TP and FN per subject and group, respectively. As seen in Figs. 3 and 4 the changes in the recall rates are observed due to the different numbers of TP and FN peaks in the controlled and free-living environments. Greater TP values are noticed in the controlled environment, which is in line with the experimental set-up when the meals are distinct. At the same time, the FN values increased under free-living conditions, where multiple, or subsequent, meal intakes were frequently observed, making it difficult to distinguish between smaller postprandial peaks and peaks that were not resulting from a meal, e.g. physical activity or natural larger fluctuations. 4.3 Algorithm Performance on the Entire Profiles (Dataset 1) Overall, the algorithm correctly identified 80% of the total number of peaks, with average recall and precision rates of 0.84 and 0.85 across all peaks in the cohort (Table 1); 20% of peaks were not detected. Differences are observed between the groups (Table 1), with the pre-DM group having the highest recall and precision rates. The precision rates across each group have a smaller range of values than the recall rates (Table 1). However, the most considerable differences are observed between individuals (Table 2 and Fig. 5) with larger differences in recall compared with the precision rate; the latter is relatively stable as seen in Fig. 5. Differences in the recall rate are more prominent within the
102
A. Archavli et al.
Fig. 4 Bar plots representing the algorithm results for each subject (a) and each group (b) for peaks recorded in free-living conditions. The blue bars represent the total number of peaks, the red bars correspond to the TP peaks and the yellow ones to the FN peaks.
no-DM and T2D groups (Table 2) due to undetected smaller peaks, which were missing by the algorithm due to the chosen threshold (stage 2 of the algorithm). The threshold was identified based on average minimum peak height values per group, making the FN peaks undistinguishable from those not caused by meals. The effect of the threshold selection on the increased number of FN peaks is evident from a T2D subject (S6, threshold = 1.3 mmol/L) with good DM management, who has the lowest recall rate of 0.56 (Table 2). Different threshold heights between 1 and 1.3 mmol/L for the same subject were additionally considered to mitigate this effect on the recall. However, this led to both the TP and FP values increasing. The increase in FP rates was greater than in TP rates resulting in a worse precision rate without a significant difference in the recall, prompting us to use the previous threshold values. Table 1. Recall and precision for each group Group
Recall
Precision
No-DM
0.75
0.86
Pre-DM
0.88
0.86
T2D
0.80
0.82
Overall
0.84
0.85
4.4 Comparison with Other Algorithms Available from Literature Table 3 compares the performance of our algorithm with other meal detection algorithms from literature developed for free-living condition data. Note that the meal detection algorithms were only concerned with identifying the starting point of peaks, whereas our algorithm aimed to detect the whole peak. The metrics used in the literature were the recall rate and the number of FP peaks per day presented as median (interquartile range) so the comparison is presented here for the same characteristics. The proposed algorithm showed an overall recall of 0.84, which is greater than that (0.7–0.79) of other published threshold-based algorithms as seen in Table 3 (Threshold
Postprandial Peak Identification from Continuous
103
Table 2. Recall and precision for each subject Group
Subject
Recall
Precision
No-DM
S1
0.77
0.94
No-DM
S2
0.63
0.88
No-DM
S3
0.83
0.80
Pre-DM
S4
0.90
0.83
Pre-DM
S5
0.85
0.89
T2D
S6
0.56
0.75
T2D
S7
0.89
0.89
T2D
S8
0.94
0.79
Fig. 5. Recall (marker o) and precision (marker +) rates of peak of detection algorithm for each indivitual
Ra, Dassau et al, Lee et al., Harvey et al., STMD). There are only three other algorithms available in the literature, which produced better results, with 0.89, 0.93 and 0.99 recall rates as seen in Table 3 (LDA Ra, LDA CGM and PAIN). However, they are not thresholdbased and invariant to individual-specific characteristics [12]. Moreover, our algorithm achieved the lowest FP value of 0.4 per day (as seen in Table 3), outperforming the rest of the algorithms. 4.5 Algorithm Evaluation on a Different Dataset (Dataset 2) An external dataset [18] containing 41 samples from individuals without DM was used for further evaluation. Glucose levels were monitored for 14 days using FreeStyle Libre (Abbott Diabetes Care, Alameda, CA) CGM, and information on physical activity, carbohydrate/meal consumption, and sleep routine (bed and wake-up times) were available
104
A. Archavli et al.
Table 3: Comparison of recall and false positive (per day presented as median (interquartile range)) rates between the proposed algorithm and other algorithms from literature [12] developed for real data. Algorithm
Recall
FP-per-day, median (iqr)
LDA Ra
0.93
1.5 (0.4)
LDA CGM
0.89
1.41 (0.42)
Threshold Ra
0.75
1.47 (0.53)
GRID
0.22
2.78 (0.41)
PAIN
0.99
1.88 (0.72)
Dassau et al.
0.74
1.62 (1.27)
Lee et al.
0.70
1.69 (1.21)
Harvey et al.
0.79
1.64 (1.34)
STMD
0.70
1.4 (1.4)
Proposed algorithm
0.84
0.4 (0)
from diaries. The first three days (days 1–3) of monitoring were excluded from the analysis, and data from days 4–14 were used for the evaluation. During days 4–5, the subjects visited the study site and received standardised, controlled meals for breakfast and lunch and a free choice dinner from a buffet at a specific time. During the next two days (days 6–7), the subjects did not visit the study site but were asked to consume three meals of their choice per day at the same times as the days in the study site and avoid carbohydrate-containing drinks and physical strain. For the rest of the monitoring period (days 8–14), the subjects followed their usual routine [18]. Days 4 and 5 were considered a controlled environment, while the rest were considered a free-living period. Out of 41 subjects, one was excluded from the analysis as they had multiple empty entries in the CGM signal. The peak detection algorithm applied to this dataset showed average recall and precision rates of 0.78 and 0.58, respectively and approximately 2 FP peaks per day (median = 2.15 (iqr = 0.95)). Recall and precision rates for each subject are presented in Fig. 6a. The precision rate dropped significantly compared with the first dataset, i.e. from 0.85 to 0.58. To examine if the results can be improved, a range of thresholds between 0.5 and 1.4 mmol/L was tested. For each threshold value, the algorithm’s results for each subject were recorded and then averaged across all the subjects (Fig. 6b). From Fig. 6b, it can be noted that smaller values of thresholds produced higher average recall rates due to a larger number of peaks being detected and lower average precision rates demonstrating that the algorithm selected a larger number of peaks that do not correspond to meals (FP) as also seen by the higher values of FP per day presented in the same figure. As the threshold grows, the recall rate decreases while the precision rate increases, and they intersect the value of 0.68 at the threshold value of 1.1. The range around this threshold value (1–1.2) could provide a better balance between recall and precision rates. To achieve a better performance, an individual threshold could also be proposed
Postprandial Peak Identification from Continuous
105
so the algorithm could adjust to individual glucose characteristics: in the cohort, 55% of individuals had a recall rate over 0.8 with a precision of 0.6.
Fig. 6 Recall (marker o) and precision (marker +) rates and FP per day (marker) of the peak detection algorithm for each individual (a) and threshold value (b). The dashed vertical line corresponds to the threshold value of 0.8 used for peak detection in the no-DM group in dataset 1.
5 Conclusions Meal/peak identification is an area of interest with multiple challenges. This work developed a local maxima threshold-based algorithm to extract the entire postprandial glucose peaks of real-life datasets and examined whether a threshold-based algorithm could be utilized for this data type from multiple glucose tolerance groups. The algorithm achieved a recall and precision of 0.84 and 0.85 respectively, with 0.4 FP per day for the dataset used in the algorithm development and involving no-DM, pre-DM and DM profiles; it achieved a recall and precision of 0.78 and 0.58, respectively, with 2.15 FP per day for a second dataset involving non-DM profiles only. Threshold-based methods require carefully chosen thresholds to achieve the appropriate sensitivity (recall), as shown from the evaluation of the algorithm with the external dataset where lower precision values were acquired. The algorithm was developed based on selected controlled food intakes. Due to the limited time the subjects spent in the controlled environment, fewer meals were considered; therefore, for certain subjects, only minimal information was available, reducing the option for a better-individualised approach as mentioned in the Sect. 4.3. Additionally, the meal timings were based on the diaries each subject kept; therefore, the margin of human error between the actual time and the stated meal time was increased, which could have affected the algorithm’s results. Achieving a perfect detection of food intake could be challenging as postpran-dial responses vary between individuals, especially between groups of different glucose tolerance. To produce better detection rates, the algorithm could be adjusted according to the purpose of peak detection. For example, when larger postprandial peaks are of interest, selecting a higher threshold would result in a better precision rate with fewer false positive peaks but more false negatives since more peaks would be excluded. However, if there is no interest in smaller peaks, greater FN values would not affect the aim of the
106
A. Archavli et al.
selection. An individualised threshold could also be derived and incorporated into the algorithm to adjust the algorithm to each profile. This approach requires insignificant computational resources and could be integrated into existing mobile phone applications synced with the CGMs.
6 Limitations of research As with the majority of studies, the design of the current study is subject to certain limitations. A limitation of the study is the small sample size in each group for the first dataset as a larger sample would provide larger variability across each group. Another limitation is the limited range and number of controlled meals in the study, not covering a variety of eating habits. Aknowledgments. The work is supported by EPSRC (UK) grant (EP/T013648/1) and the University Hospitals Coventry and Warwickshire (UK).
References 1. Saeedi, P., et al.: Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: Results from the International Diabetes Federation Diabetes Atlas, 9th edn. Diabetes Res. Clin. Pract. 157, 107843 (2019). https://doi.org/10.1016/j.diabres.2019. 107843 2. Sun, H., et al.: The status and trends of diabetes in China: a systematic review and metaanalysis. Diabetes Res. Clin. Pract. 183, 109119 (2022). https://doi.org/10.1016/j.diabres. 2021.109119 3. Gillani, S.W., et al.: Predictors of health-related quality of life among patients with type ii diabetes mellitus who are insulin users: a multidimensional model. Curr. Ther. Res. Clin. Exp. 90, 53–60 (2019). https://doi.org/10.1016/j.curtheres.2019.04.001 4. Aleppo, G., Webb, K.: Continuous glucose monitoring integration in clinical practice: a stepped guide to data review and interpretation. J. Diabetes Sci. Technol. 13(4), 664–673 (2018). https://doi.org/10.1177/1932296818813581 5. Beck R.W., Bergenstal R.M.: Continuous glucose monitoring for type 2 diabetes: how does it compare with type 1 diabetes? Diabetes Technol. Ther. 24(3), 153–156 (2022). https://doi. org/10.1089/dia.2021.0374 6. Ahn, Y.C., et al.: Effectiveness of non-contact dietary coaching in adults with diabetes or prediabetes using a continuous glucose monitoring device: a randomized controlled trial. Healthcare 11(2), 252 (2023). https://doi.org/10.3390/healthcare11020252 7. Holzer, R., Bloch, W., Brinkmann, C.: Continuous glucose monitoring in healthy adults— possible applications in health care, wellness, and sports. Sensors 22(5), 2030 (2022). https:// doi.org/10.3390/s22052030 8. Fico, G., et al.: Exploring the frequency domain of continuous glucose monitoring signals to improve characterization of glucose variability and of diabetic profiles. J. Diabetes Sci. Technol. 11(4), 773–779 (2017). https://doi.org/10.1177/1932296816685717 9. Zheng, M., Ni, B., Kleinberg, S.: Discriminating power: a privacy-preserving distributed algorithm for learning decision trees. J. Am. Med. Inform. Assoc. 26(12), 1592–1599 (2019). https://doi.org/10.1093/jamia/ocz159
Postprandial Peak Identification from Continuous
107
10. Samadi, S., et al.: Meal detection and carbohydrate estimation using continuous glucose sensor data. IEEE J. Biomed. Health Inform. 21(3), 619–627 (2017). https://doi.org/10.1109/ JBHI.2017.2677953 11. Daniels, J., Herrero, P., Georgiou, P.: A deep learning framework for automatic meal detection and estimation in artificial pancreas systems. Sensors 22(2), 466 (2022). https://doi.org/10. 3390/s22020466. 12. Faccioli, S., et al.: Super–twisting-based meal detector for type 1 diabetes management: Improvement and assessment in a real-life scenario. Comput. Methods Programs Biomed. 219, 106736 (2022). https://doi.org/10.1016/j.cmpb.2022.106736 13. Palacios, V., et al.: Machine learning based meal detection using continuous glucose monitoring on healthy participants: an objective measure of participant compliance to protocol. Conf. Proc. IEEE Eng. Med. Biol. Soc. 7032-7035 (2021). https://doi.org/10.1109/EMBC46 164.2021.9630408 14. Eichenlaub, M.M.W.: Mathematical modelling of blood glucose dynamics in normal and impaired glucose tolerance. Ph.D. thesis. University of Warwick (2020) 15. Zhang, Y., Holt, T.A., Khovanova, N.: A data driven nonlinear stochastic model for bloodglucose dynamics. Comput. Methods Programs Biomed. 125, 18–25 (2016). https://doi.org/ 10.1016/j.cmpb.2015.10.021 16. ElSayed N.A., et al.: 6. Glycemic targets: standards of care in diabetes-2023. Diabetes Care 46(Suppl 1), S97–S110 (2023). https://doi.org/10.2337/dc23-S006 17. Cheng, X., et al.: The shape of the glucose response curve during an oral glucose tolerance test heralds β–cell function in a large Chinese population. BMC. Endocr. Disord. 19(1), 119 (2019). https://doi.org/10.1186/s12902-019-0446-4 18. Freckmann, G., et al.: Continuous glucose profiles in healthy people with fixed meal times and under everyday life conditions. J. Diabetes Sci. Technol. 19322968221113341 (2022). https://doi.org/10.1177/19322968221113341
Emotional State Evaluation in Driving Simulations: PC Versus Virtual Theater Rita Laureanti(B) , Simone Mentasti, Alessandro Gabrielli, Matteo Matteucci, and Luca Mainardi Department of Electronics, Information and Bioengineering (DEIB), Politecnico di Milano, Milan, Italy [email protected]
Abstract. Emotional assessment during driving is a developing research field. In this study, two set-ups of driving simulation were tested: a laboratory set and a virtual theater set. People emotional state was evaluated in the two settings through questionnaires and continuous recording of physiological signals, such as electrodermal activity and heart rate. Contrasting results were found considering the declared by the participants (virtual theater preferred, more engaging, less stressful) and the information obtained by the physiological response (higher positive emotions in the laboratory). We also found higher skin conductance and heart rate for the subgroup with no experience of driving games and slightly higher sympathetic activation in the group with 3 variables [9]. The O-information framework was further advanced with the introduction of the O-Information rate, which offers a dynamic representation of synergy and redundancy that was extended also in the frequency domain [4]. Most of the existing measures generally do not consider the interaction be tween pairs of nodes in the context of the entire network. To address this aspect, the present paper expands the concept of “local O-information” from [9] and introduces a new “balance index” (B-index) that considers the balance between synergy and redundancy in pairwise functional connectivity extended to HOIs. Combining measures of Mutual Information (MI) and Conditional Mutual In formation (CMI), the index provides a robust and reliable representation of the network structure by utilizing pairwise links. Moreover, it is implemented to reduce the incidence of false outcomes caused by common driver, chain, or com mon child effects [10], and therefore to offer a more accurate understanding of the network’s functional structure. In this work, the B-index is initially validated on a simulated network of in teracting random variables, demonstrating its ability to accurately reconstruct the actual direct links between nodes and their synergistic or redundant behavior in relation to the rest of the network. Afterwards, it is applied in two practical cases: functional magnetic resonance imaging (fMRI) of a pediatric patient with hepatic cavernoma at different post-operation recovery stages, and in-vitro cultured cortical neuronal cells at various time points of neuronal development.
Synergy and Redundancy
2
147
Methods
Let S = {S1, · · · , SM } be a network of M independent and identically distributed (i.i.d.) random variables. This work considers two variables X = Si and Y = Sj , where i, j = 1, · · · , M and i = j, and the remaining M − 2 variables collected in the vector Z = S \ [X, Y ], adopting a static-analysis approach that disregards temporal correlations within each process and only examines zero-lag correlations between the variables X, Y and Z. 2.1
Formulation of the B-Index
The relationships between the variables in S can be analyzed by using well-known measures of mutual information (MI) and conditional MI (CMI). Given the random variables X, Y and the random vector Z taken from S, the MI measures the amount of information shared between X and Y, while CMI evaluates the information shared between X and Y when considering the influence of the third variable Z. The MI and CMI are defined respectively as I(X; Y ) = H(X)−H(X | Y ), I(X; Y | Z) = H(X | Z) − H(X | Y, Z), where H(·) and H(·|·) denote entropy and conditional entropy [11]. Our formulations rely on the fact that the II between X, Y , and Z can be described as the difference between the MI I(X; Y ) and the CMI I(X; Y | Z), which reflects the balance between redundancy and synergy and is referred to the local O-Information [9]. To quantify this balance, we define B-index as a normalized form of interaction information: B(X; Y ) =
I(X; Y ) − I(X; Y | Z) max{I(X; Y ), I(X; Y | Z)}
(1)
Normalizing the difference between MI and CMI by their maximum enables to highlight the redundant or synergistic nature of the interaction and, at the same time, to uncover the underlying connectivity structure of the network. In fact, the value of B(X; Y) can be either positive or negative and is bounded between -1 and 1. Specifically, the B-index takes the following values: • B = 1, denoting maximum redundancy which occurs when a false positive link is detected due to the effects of a common driver or cascade (MI > 0, CMI = 0); • B = −1, denoting maximum synergy when a false positive link is detected due to the effect of a common child (MI = 0, CMI > 0); • −1 < B < 1, denoting prevalence of redundancy when B > 0(MI > CMI) and prevalence of synergy when B < 0(CMI > MI) and reflecting the presence of a true positive link; • B = 0, denoting a full balance between redundancy and synergy (MI = CMI > 0) and reflecting the presence of a true positive link between two nodes isolated from the rest of the network; • B = NaN, denoting the absence of synergy and redundancy in disconnected nodes (MI = CMI = 0) and reflecting a true negative link.
148
Y. Antonacci et al.
To accurately reconstruct the network structure based on B-index, it is crucial to assess the statistical significance of MI and CMI measures through techniques such as surrogate data analysis. This ensures that the B-index accurately reflects fully redundant or synergistic interactions with values of 1 or −1 respectively, and identifies the absence of interactions with a return of NaN. The procedure used in this work to generate surrogate data is described in the next section. 2.2
Computation of B-Index
Here we report the different approaches for the numerical computation of the B index for both continuous random variables (assuming a Gaussian distribution) and discrete variables (assuming binary states). a. Gaussian Variables Under the assumption that X, Y and Z have a joint Gaussian distribution, the MI and CMI can be assessed from the following linear regression models: X = aY + U, X = bZ + W, X = cY + dZ + V,
(2)
whereby the variable X is first predicted from the variable Y weighed by the coefficient a, then from the vector Z weighed by the vector of coefficients b, and finally from the variable Y and the vector Z respectively weighed by the coefficient c and the vector of coefficients d; U, W and V are zero-mean i.i.d. variables modelling the prediction errors. MI and CMI are assessed exploiting the relations between entropy and variance and between conditional entropy 2 ), and partial variance valid for Gaussian variables [12], i.e. H(X) = 12 ln (2πeσX 1 2 2 H(X|Y ) = 2 ln (2πeσU ), H(X|Z) = ln (2πeσW ), and H(X|Y, Z) = ln (2πeσV2 ), 2 2 2 is the variance of X, while σU , σW and σV2 are the variances of the where σX residuals U , W and V [13]. This allows to formulate the MI and CMI as: 2 σX 1 I(X; Y ) = ln 2 σ2 (3) 2U σW 1 I(X; Y | Z) = ln 2 σV2 b. Binary Variables When considering discrete random variables X, Y, and each component of the vector Z that take values from the binary alphabet A ∈ {0, 1}, the MI and CMI are defined, respectively, as follows: p(x, y) , p(x, y) log2 I(X; Y ) = p(x)p(y) x,y∈{0,1} (4) p(x, y | z) I(X; Y | Z) = , p(x, y, z) log2 p(x | z)p(y | z) x,y,z∈{0,1}
Synergy and Redundancy
149
where p(·), p(·, ·), and p(· | ·) denote the marginal, joint, and conditional proba bility density, respectively, which are determined by calculating the proportion of occurrences of “0” and “1” in the observed binary sequences. The method of surrogate data was used to assess the statistical significance of both MI and CMI. Specifically, for each node in the observed network, 100 surrogates were generated through a random shuffling method, and MI and CMI were calculated between each pair of surrogates. To assess the significance of the original MI and CMI values, a 5% significance threshold was established based on the 95 apex percentile of the distribution of the surrogate MI and CMI values. Any original MI and CMI values below the corresponding thresholds were deemed non-significant, and were set to zero to ensure values of -1, 1 or NaN for the B-Index (Eq. 1).
3
Simulation Study
In this section, we study the behavior of the B-index by using a simulated linear model in the form S = AS + E, where S = [S1 , · · · , SM ]T is a vector of M = 10 random variables, E = [E1 , · · · , EM ]T is a vector of M independent noise terms, and A is a M × M matrix of connectivity coefficients representing a directed acyclic graph for which self-loops are not allowed. If Ai,j = 0, then the model implies the direct interaction Sj → Si , otherwise Si and Sj are not directly connected. To simulate data for this liner model, the above equation can be rewritten as S = [I − A]−1 E, where I is the M × M identity matrix. The vector variable textbf S is then given N = 1000 data-points, with the coefficients of the connectivity matrix A specified according to the structure and values of Ai,j shown in Fig. 1a. The noise terms in the data are sampled from a standard normal distribution with zero mean and unit variance. The redundant or synergistic behavior of each link in relation to the rest of the simulated network, as well as the reconstruction of the network structure, are illustrated respectively in Fig. 1b and c.
Fig. 1. a) Schematic representation of the imposed connectivity pattern, in which each connection is assigned a weight of Ai,j = 1.5. b) Matrix of B-index values obtained for each pair of nodes in the simulated network. c) Network structure assessed using the B index, with links colored in red or blue to indicate redundancy or synergy, respectively, while white represents a state of perfect balance between the two
150
Y. Antonacci et al.
In the following, we explore particular instances of simulated spurious effects and the related behavior of the B-index: (1) As a first example, it can be seen that nodes 6 and 7 are unconnected, but share a common cause in node 5(6 ← 5 → 7). This results in a B-value of 1, indicating maximum redundancy (Fig. 1b). The use of MI would result in a false positive link, which is correctly not identified when using CMI. Since only MI is statistically significant, the B-structure correctly shows no connection between nodes 6 and 7 in Fig. 1c. (2) In another example, we consider a chain structure (5 → 7 → 8). The B-value of the connection between nodes 5 and 8 is equal to 1 (Fig. 1b), indicating a high degree of redundancy and no visible connection in the B-structure (Fig. 1c). Similar to the previous example, while MI is statistically significant, CMI is not. This is because CMI is able to distinguish between direct and indirect interactions in the presence of chains, avoiding false-link detection. (3) The limitation of CMI becomes apparent In the presence of common child structures (e.g. 3 → 2 ← 4, or 4 → 2 ← 5). In these cases, 2 is a common effect of 3 and 4 (or 4 and 5), which are otherwise independent. The non-zero CMI between 3 and 4 (or 4 and 5) may result in false link detection if MI is not computed. The B-index value of -1 in this scenario indicates maximum synergy due to statistically significant MI only, resulting in no connection between 3 and 4 (or 4 and 5) in the B-structure. Figure 1c). (4) When nodes are connected but isolated, as with nodes 9 and 10, the B-index is equal to zero, indicating a perfect balance between redundancy and synergy. When the B-index falls between −1 and 1, it provides insight into the presence or absence of synergistic behavior of that specific link.
4
Application to the Real Data
To investigate static connectivity in experimental recordings by using the pro posed B-index, we analyzed two sources of data. The first source involved blood oxygenation level dependent (BOLD) signals obtained from resting-state func tional magnetic resonance imaging (rest-fMRI) in a pediatric patient with hep atic cavernoma, where the B-index was calculated based on MI and CMI under the assumption of a Gaussian distribution (Eq. 3). In the second application, we analyzed a single in-vitro developing culture of dissociated cortical neurons, where B-index was calculated based on MI and CMI defined for discrete random variables (Eq. 4), which were obtained by transforming observed neural firing activity into binary streams with discrete states “0” and “1”. 4.1
fMRI Data
The analyzed patient was affected by hepatic cavernoma, which resulted in a por tosystemic shunt and encephalopathy, and subsequently underwent liver vascular shunt correction. The rest-fMRI with BOLD echo-planar imaging technique was performed to assess the spontaneous neuronal activity within the resting-state networks, before (PRE), after 6 months (POST1) and 12 months (POST12) the surgical correction of the portosystemic shunt. The CONN toolbox was used to
Synergy and Redundancy
151
Fig. 2. The B-index estimates (a) and the reconstructed network structures (b) com puted across the resting-state networks of 32 BOLD fMRI series representing the pre surgery state (PRE), 6 months post-surgery (POST1) and 12 months post-surgery (POST12) due to an hepatic cavernoma issue in a pediatric patient
perform the seed-based extraction of M = 32 BOLD series as sequences of N = 200 consecutive synchronous values, which are then considered as a realization of the vector process S = {S1 , . . . , S32 } describing the neural dynamics. From each combination of BOLD time series, pre-processed by removing the mean value, the analysis was performed for each experimental condition by com puting the MI and the CMI between them, assessing their statistical significance with surrogate data and evaluating the B-index. Results of the analysis are reported in Fig. 2, showing the subject-specific maps of brain connectivity (Fig. 2, left) and the reconstructed network struc tures (Fig. 2, right), in the three experimental conditions. Compared to the pre-surgery phase (PRE), the immediate post-surgery period (POST1) is char acterized by a weakening of brain connectivity. This is demonstrated by lower density of both redundant and synergistic links in the B structure (Fig. 2b). Interestingly, twelve months after the treatment (POST12), the brain network is mostly synergistic, with increased number of true connections demonstrated by an augmented B density (Fig. 2c). In this patient the proposed surgery correction of the portosystemic shunt worked in recovering brain cognition, which is associated by our analysis with stronger brain connectivity and an increase of synergistic interactions between the nodes of the investigated resting state networks.
152
4.2
Y. Antonacci et al.
In-Vitro Neuronal Culture
For the second application, we selected a high-density in-vitro culture (2500 cells/µL) from a public repository of dissociated neuron cultures at different stages of development, measured by days in-vitro (DIV) [14]. The culture was monitored during a day and by using a multi-electrode array (MEA) with 59 elec trodes, where each electrode recorded the collective spiking activity of roughly 100 to 1000 neurons (multi-unit activity, MUA). We observed the MUA of the selected culture at three stages of neuronal development: early (7 DIV), inter mediate (15 DIV), and mature (25 DIV). Given the tendency of these neuronal cells to self-arrange into bursting activity, our analysis was focused on this phe nomenon. Moreover, each MUA was transformed into a discrete binary stream, where state 1 coarsely represents the occurrence of bursting activity, and state 0 represents inter-burst periods. This was done by following the approach proposed in [15], after initial burst identification as done in [16]. Finally, the B-index was calculated over a vector of 59 binary profiles S = {S1 , · · · , S59 }, for each stage of neuronal development. Figure 3 presents the result of estimation of the B-index between all pairs of binary profiles and the corresponding reconstructed B-structures. Our findings demonstrate significantly reduced both network activity and connectivity dur ing earlier stages of neural development, which is in line with previous studies [16]. Additionally, we found that connections in earlier stages when existed were predominantly influenced by spurious correlations, resulting in the B-index and
Fig. 3. The B-index estimates (a) and the reconstructed network structures (b) com puted across networks of 59 electrodes representing the activity of a randomly selected in-vitro neuronal culture at various stages of neuronal maturation (early, ∼ 7 DIV; intermediate, ∼ 15 DIV; mature, ∼ 25 DIV)
Synergy and Redundancy
153
structure matrices being almost void at the DIV6 (Fig. 3a) and DIV15 (Fig. 3b) stages. Only in the mature stage (DIV25) the true links were identified, and bursting patterns exhibited purely functionally-redundant interactions among electrodes (Fig. 3c). This preliminary finding indicates that redundant connec tivity may play a crucial role in the robustness and flexibility of neural networks even in in-vitro cultures.
5
Conclusion
The proposed B-index constitutes a reliable method for determining the func tional structure of a network, even when confounding factors are present. In the simulated network, we demonstrated that our approach reduces the occurrence of false positives and facilitates the distinction between direct and indirect inter actions. Moreover, the connection to local O-Information yields valuable insights into the synergistic or redundant interaction of each link with the rest of the net work. Thus, the combination of network reconstruction and evaluation of HOIs makes our approach useful across different domains, including neuroscience and physiology. In future works, the focus will be on incorporating more real-data subjects and applications in the context of Network Physiology [17,18], and on refining the index definition by using mutual information rates instead of just mutual information, which will make the measure dynamic and able to account for temporal correlations. Acknowledgement. This study was supported by SiciliAn MicronanOTecH Research And Innova tion CEnter “SAMOTHRACE” (MUR, PNRR-M4C2, ECS 00000022), spoke 3 - Universita’ degli Studi di Palermo“S2-COMMs - Micro and Nanotechnologies for Smart & Sustainable Communities.
References 1. Battiston, F., Cencetti, G., Iacopini, I., Latora, V., Lucas, M., Patania, A., Young, J.-G., Petri, G.: Networks beyond pairwise interactions: structure and dynamics. Phys. Rep. 874, 1–92 (2020) 2. Martinez-Gutierrez, E., Jimenez-Marin, A., Stramaglia, S., Cortes, J.M.: The structure of anticorrelated networks in the human brain. Front. Netw. Physiol. 53 (2022) 3. G¨ unther, M., Kantelhardt, J.W., Bartsch, R.P.: The reconstruction of causal networks in physiology. Front. Netw. Physiol. 2 (2022) 4. Faes, L., Mijatovic, G., Antonacci, Y., Pernice, R., Bar, C., Sparacino, L., Sammartino, M., Porta, A., Marinazzo, D., Stramaglia, S.: A new framework for the time-and frequency-domain assessment of high-order interactions in networks of random processes. IEEE Trans. Signal Process. 70, 5766–5777 (2022) 5. Rosas, F.E., Mediano, P.A., Luppi, A.I., Varley, T.F., Lizier, J.T., Stramaglia, S., Jensen, H.J., Marinazzo, D.: Disentangling high-order mechanisms and high order behaviours in complex systems. Nat. Phys. 18(5), 476–477 (2022) 6. McGill, W.: Multivariate information transmission. Trans. IRE Prof. Group Inform. Theory 4(4), 93–111 (1954)
154
Y. Antonacci et al.
7. Williams, P.L., Beer, R.D.: Nonnegative decomposition of multivariate information. arXiv preprint arXiv:1004.2515 (2010) 8. Lizier, J.T., Bertschinger, N., Jost, J., Wibral, M.: Information decomposition of target effects from multi-source interactions: perspectives on previous, current and future work, p. 307 (2018) 9. Rosas, F.E., Mediano, P.A., Gastpar, M., Jensen, H.J.: Quantifying high-order interdependencies via multivariate extensions of the mutual information. Phys. Rev. E 100(3), 032305 (2019) 10. Sanchez-Romero, R., Cole, M.W.: Combining multiple functional connectivity methods to improve causal inferences. J. Cogn. Neurosci. 33(2), 180–194 (2021) 11. Cover Thomas, M., Thomas Joy, A.: Elements of Information Theory, vol. 3, pp. 37–38. Wiley, New York (1991) 12. Barnett, L., Barrett, A.B., Seth, A.K.: Granger causality and transfer entropy are equivalent for gaussian variables. Phys. Rev. Lett. 103(23), 238701 (2009) 13. Barrett, A.B., Barnett, L., Seth, A.K.: Multivariate granger causality and generalized variance. Phys. Rev. E 81(4), 041907 (2010) 14. Wagenaar, D.A., Pine, J., Potter, S.M.: An extremely rich repertoire of bursting patterns during the development of cortical cultures. BMC Neurosci. 7(1), 1–18 (2006) 15. Mijatovi´c, G., Lonˇcar-Turukalo, T., Procyk, E., Baji´c, D.: A novel approach to probabilistic characterisation of neural firing patterns. Journal of neuroscience methods 305, 67–81 (2018) 16. Mijatovic, G., Antonacci, Y., Loncar-Turukalo, T., Minati, L., Faes, L.: An information-theoretic framework to measure the dynamic interaction between neural spike trains. IEEE Trans. Biomed. Eng. 68(12), 3471–3481 (2021) 17. Schmal, C., Hong, S., Tokuda, I., Myung, J.: Coupling in biological systems: definitions, mechanisms, and implications (2022) 18. Ivanov, P.C.: The new field of network physiology: building the human physiolome. Front. Netw. Physiol. 1 (2021)
Comparison of Linear Model-Based and Nonlinear Model-Free Directional Coupling Measures: Analysis of Cardiovascular and Cardiorespiratory Interactions at Rest and During Physiological Stress Chiara Barà1(B) , Riccardo Pernice1 , Laura Sparacino1 , Yuri Antonacci1 , Michal Javorka2 , and Luca Faes1 1 Department of Engineering, University of Palermo, Palermo, Italy
[email protected] 2 Department of Physiology and Biomedical Center Martin Jessenius Faculty of Medicine,
Comenius University, Martin, Slovakia
Abstract. In this work, we present an investigation of the cardiovascular and cardiorespiratory regulatory mechanisms involved during stress responses using the information-theoretic measure of transfer entropy (TE). Specifically, the aim of the study is to compare different estimation approaches for the evaluation of the information transferred among different physiological systems. The analysis was carried out on the series of heart period, systolic arterial pressure and respiration measured from 61 young healthy subjects, at rest and during orthostatic and mental stress states, by using both a linear model-based and a nonlinear modelfree approaches to compute TE. The results reveal mostly significant correlations for the measures of TE estimated with the two approaches, particularly when assessing the influence of respiration on cardiovascular activity during mental stress and the influence of vascular dynamics on cardiac activity during tilt. Therefore, our findings suggest that the simpler linear parametric approach is suitable in conditions predominantly regulated by sympathetic nervous system or by the withdrawal of the parasympathetic system. Keywords: Multivariate time series analysis · Transfer entropy · Cardiovascular system · Respiration · Stress
1 Introduction The human body is a complex network of interdependent systems and control mechanisms that work together to maintain homeostasis [1]. However, currently there is not well-established analytical methodologies, computational tools, and theoretical frameworks that can effectively extract and quantify the interactions between physiological systems from continuous streams of data. The dynamics of these interactions are complex and occur across multiple scales, which presents significant challenges in identifying © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 155–163, 2024. https://doi.org/10.1007/978-3-031-49062-0_17
156
C. Barà et al.
and measuring these networks. While traditional methods such as cross-correlation and coherence have been proposed to investigate the relationships between two systems [2], there is growing interest in the use of the framework of information-theory to better understand the complex interplay between organ systems [3]. Turing’s theory suggests that any type of information carried by a system can be broken down into three distinct components: information storage, information transfer, and information modification [4]. The measurement of information transfer is commonly achieved through the use of transfer entropy (TE), which calculates the directional effects between two processes by analyzing the information provided by a putative driver system above and beyond the predictability given by the target itself. Related to the concept of Granger causality [5], the TE measure has been shown to be a valuable tool for assessing information transfer between interconnected systems in various contexts [5–9]. A main challenge regards the practical estimation of the TE from short realizations of physiological processes [10]. Over the years, various methodologies have been proposed in the literature for estimating transfer entropy, which can be classified as either model-based or model-free. Both approaches rely on the assumption of stationarity, but in the former case, the probability density function used to calculate TE can be fully described by a specific model, while in the latter case, it is directly estimated from the data [11]. When dealing with a linear model, the model-based approach requires less data and computational resources and should be chosen when only short datasets are available. However, the model-free method is crucial for revealing the intricate structure of physiological connections supported by nonlinear feedback interactions [12, 13]. In this context, the present study proposes an analysis of the information transfer assessed from short term series reflecting the spontaneous variability of heart period, systolic arterial pressure and respiration in healthy subjects undergoing a protocol including orthostatic and mental stress. In particular, the main goal is to compare two different estimation approaches, i.e., a simpler linear parametric and a more sophisticated non linear k-nearest neighbors method.
2 Material and Methods 2.1 Subjects, Experimental Protocol and Time Series Extraction In this study, an historical database previously employed to study cardiovascular variability was used [9, 14]. In detail, electrocardiographic (ECG; CardioFax ECG-9620, NihonKohden, Japan), arterial blood pressure (ABP; Finometer Pro, FMS, Netherlands) and respiratory volume (RespiTrace, NIMS, USA) signals acquired at a sampling rate of 1 kHz on a group of sixty-one healthy young subjects (37 females, 17.5 2.4 years old) were analysed [9]. The experimental protocol was approved by the Ethical Committee of the Jessenius Faculty of Medicine, Comenius University, Martin, Slovakia, and consisted of (i) a supine resting state (REST, 15 min), (ii) an orthostatic stress phase (HUT, 8 min) during which an head-up tilt test was performed tilting a motorized table to 45 degrees, and (iii) a mental arithmetic stress phase (MA, 6 min) during which subjects were instructed to perform a series of additions with three-digit numbers until a one-digit number was reached, and to indicate whether the result was odd or even. Both HUT and
Comparison of Linear Model-Based and Nonlinear Model-Free Measures
157
MA phases were followed by 10 min of supine recovery, which were not considered in our analysis. From these signals, the heart periods (RR) time series was measured as the time differences between two consecutive ECG R peaks, the systolic arterial pressure (SAP) time series was obtained by identifying the maximum values of the ABP signal within each detected heart period, while the respiratory (RESP) one by sampling the respiratory volume signal at each detected ECG R peak. Time series windows of 300 samples were extracted for each subject and protocol phase discarding any transient periods, as detailed in [9] to which we refer the reader for further information about data acquisition and experimental protocol. Before computing the directional coupling measures, all time series were preprocessed applying a zero phase high-pass autoregressive filter, by removing and interpolating samples differing more than three times standard deviation from the mean value, and normalizing the series to zero mean and unit variance. 2.2 Directional Coupling Measure Given a bivariate system S = {X , Y}, its joint and dynamical evolution can be described looking at the stochastic processes X and Y. Let us indicate as X n , Y n the scalar random variables describing the current state of X and Y, as X kn = [X n−1 · · · X n−k ] ∈ Rk × 1 , Y kn = [Y n−1 · · · Y n−k ] ∈ Rk × 1 the vector variables sampling the two processes over the past k lags, and as X 1n = limk→∞ X 1n =, Y kn = limk→∞ Y kn the infinitedimensional variables sampling X and Y over their whole past history. Considering X and Y respectively as driver and target processes, the information transferred from X to Y is defined as [7]: − − − − TEX →Y = I (Yn ; X − n |Y n ) = H (Yn |Y n ) − H (Yn |X n , Y n ),
(1)
where I(·;·|·) and H(·|·) denote conditional mutual information and conditional entropy, respectively. The TE measure reflects the directional influence of the driver process X on the target one Y; in absence of any interaction between the dynamical systems X and Y, the whole predictive information about the target is stored in its own past history and the information transferred from X to Y is equal to zero. Given a pair of time series {x, y} representing a realization of the driver and target processes, TE estimates can be obtained either from the parametric representation of the bivariate system dynamics, using a model-based (MB) approach, or directly from the probability density distribution of data, using a model-free (MF) approach. In the follow, the two approaches are described considering two joint Markov processes of p p×1 and order p, whose past processes can be approximated to p lags, i.e., X − n ≈ Xn ∈ R p − p×1 Yn ≈ Yn ∈ R . Linear parametric method. Under the hypothesis of gaussianity of the investigated stochastic processes, the current state of the target process Y n can be described in terms p of the past history of the driver process X n by using the autoregressive (AR) model, i.e., p p p Yn = AY n + Un , and of both the driver and the target processes [X n , Y n ] via the crossp p AR (ARX) model, i.e., Yn = AY n + BX n + Wn , where A and B are the vectors of the models coefficients belonging to the space R1×p , while U n and W n are two zero-mean
158
C. Barà et al.
2 , respectively. In this framework, the linear white Gaussian noises of variance σU2 and σW estimate (lin) of the information transferred from the X process to Y one is expressed as:
TElin =
1 σU2 ln 2 , 2 σW
(2)
where the variance of the prediction error of the AR and ARX models, i.e., σU2 and 2 , can be obtained through the identification of the two models via the Ordinary Least σW Square (OLS) method. Nonlinear model-free method. The model-free estimation of the information transferred from X to Y was performed using the Kraskov-Stögbauer-Grassberger formulation of the k-nearest neighbors (knn) approach [15]. Based on a neighbour search of samples in the highest dimensional space and then on range searches in the projected lower dimensional spaces for estimating the probability data distribution, TE is computed as [6]: (3) TEknn = ψ(k) + ψ NY pn + 1 − ψ NYn Y pn + 1 − ψ NX pn Y pn + 1 , where ψ(·) is the digamma function, k is the number of neighbors, n,k is twice the distance from each point to its kth neighbor in the higher dimensional space p p (i.e., [Yn Y n X n ]), and N Z is the number of neighbors whose distance in the Z space is lower than n,k / 2. 2.3 Data and Statistical Analysis The two formulations of transfer entropy described in Sect. 2.2 were used to compute the information transferred from X to Y, being X = RESP and Y = RR when considering the cardiorespiratory system, X = RESP and Y = SAP when considering vascular and respiratory systems, and X = SAP and Y = RR when considering the cardiovascular system. The MB approach was implemented using the Bayesian Information Criterion (BIC) to set the optimal orders p for AR and ARX models (with the maximum order was fixed to 10) [16], while the MF approach was implemented through a non-uniform embedding technique which minimises the conditional mutual information [6]. According to previous works [6, 17], the number of neighbors k was fixed to 10, while a maximum lag of 10 and 100 random shuffling surrogates have been used to establish the exit criterion. Non-parametric statistical tests were used for each estimator to identify differences in the TE among physiological states, since the assumption of gaussianity of the transfer entropy distributions was rejected using the Anderson-Darling test. Specifically, the Kruskal-Wallis analysis of variance followed by the Wilcoxon post-hoc signed rank test with Bonferroni correction (n = 3) was applied. Moreover, the statistical significance of the estimated TE value was assessed for each subject, condition and estimator using 100 surrogates generated randomly shifting the target series over time (minimum shift of 20 lags). The McNemar test for paired proportions was then carried out to determine significant variations between conditions of the number of subjects showing significant TE.
Comparison of Linear Model-Based and Nonlinear Model-Free Measures
159
The Pearson product-moment correlation coefficient r was computed between the TE measures computed through MB and MF approaches for each direction and protocol phase, testing the null hypothesis of absence of linear relation between the two estimates of information transfer. For all statistical tests, the significance level was set at 0.05. Data processing and analysis were performed using MATLAB 2020a (The Mathworks, Inc.); codes used to estimate transfer entropy measures are collected in the ITS toolbox (http://www.lucafaes.net/its.html).
3 Results and Discussion Figure 1 reports the distributions across subjects of the information transfer from RESP to RR and SAP, and from SAP to RR, during the three physiological states (i.e., rest, orthostatic and mental stress). A decrease of the influence of the ventilation activity on both cardiac (Fig. 1a) and vascular (Fig. 1b) dynamics is reported during physical (in pink) and mental stress (in green) states if compared with resting (in light blue). This decrease is associated with a reduction in the activity of the parasympathetic nervous system (PNS) [18, 19] evoked by stress, which leads to a weakened respiratory sinus arrhythmia (RSA) mechanism [9, 20], as well as a decrease in the mechanical effect of ventilation during mental arithmetic [21]. Moreover, the causal influence of SAP on RR increases significantly during head-up tilt, but not during mental arithmetics, compared to the resting condition (Fig. 1c). This results is related to the fact that tilting activates the feedback mechanism in the closed-loop cardiovascular regulatory system [18]. Previous studies have demonstrated that ventilation activity is a primary source of cardiovascular variability related to stress [2, 9], as here evidenced by the results of both the surrogate data analysis (Fig. 1d–f), with the significance of TE SAP →RR being lower than the other two measures, and the McNemar test. The reported differences are detected statistically significant using both estimators, except for the decrease in directional coupling from RESP to RR during tilt, which is only detected by the nonlinear estimator (Fig. 1a). The agreement between linear and knn estimates of the TE is corroborated by the correlation analysis results presented in Fig. 2. The correlation between the parametric and the model-free estimates of TE RESP →RR is low during REST (Fig. 2a) and increases with stress, with the maximum value reported during MA and the highest significance as well (Fig. 2g). This finding can be explained by the linearization of cardiorespiratory dynamics during stress, as nonlinearities are known to be common in healthy individuals in situations with a dominant vagal modulation [11, 13, 22]. When looking at the influence of the respiratory activity to the vascular system, we may infer that nonlinear dynamics are predominant during HUT, as evidenced by a lower correlation of the measure TE RESP →SAP during tilting (Fig. 2e) compared to the other two conditions (Fig. 2b, h). As reported in literature, this can be related to the nonlinear influence of breathing activity on cardiac baroreflex and venous return mechanisms [23]. On the other hand, when investigating the cardiovascular regulatory mechanism, a nearly constant correlation between parametric and model-free estimates is observed throughout the various phases of the protocol (Fig. 2c, f, i). Indeed, previous
160
C. Barà et al. (a) 0.3
(b)
*
* *
[nats]
0.4
* *
0
60
* *
* *
R T M lin
R T M knn
0.2 0.15
0.2
0.1
0.2
0.1
no subjects
(c)
* *
0.05
R T M lin (d) 59 59 59
R T M knn 59
0
60
53
57
R T M lin (e)
R T M knn
61 61 61
61 61 60
0
60
(f)
57
40
40
40
20
20
20
0
0
0
49
*
**
49
15
R T M lin
R T M knn
R T M lin
R T M knn
*
39
R T M lin
20
R T M knn
Fig. 1. Error bar graphs of TE computed (a) from RESP to RR, (b) from RESP to SAP and (c) from SAP to RR for all subjects during rest (R, light blue), orthostatic stress (T, pink) and mental stress (M, green) states with both linear (lin) and nonlinear (knn) estimators. In panels a–c, the asterisks indicate a statistically significant difference comparing the given phase to the one corresponding to the asterisk colour according to Wilcoxon test with Bonferroni correction (n = 3). The KruskalWallis analysis results are not reported in the figure, being always statistically significant. The number of subjects with statistically significant TE according to surrogate analysis are reported in the bar plots in panels d–f, and the asterisks indicate a statistically significant difference comparing conditions according to McNemar test.
studies have provided evidences that sympathetic activity predominantly regulates SAP dynamics and that such activity is not reflected in nonlinear dynamics [24]. As regard the comparison of the two estimation approaches in discriminating the two stress conditions, similar findings have been found on heart rate variability during tilt using the univariate measure of conditional entropy [11]. Furthermore, our results are supported by previous researches on the interaction between different physiological systems during physical or mental stress, which reported comparable trends in the different directional measures using both linear [9] and nonlinear [6] approaches.
Comparison of Linear Model-Based and Nonlinear Model-Free Measures
REST knn [nats]
0.6
(a)
0.3
r = 0.52 p = 2.13e-05
0.4
0.2
0.2
0.2
0.1
(d)
HUT knn [nats]
(b)
0.4
0
0
0.1 0.2 0.3
0.6
0.4 r = 0.58 p = 1.09e-06 0.3
r = 0.63 p = 5.67e-08
0.02 0.06 0.1 0.14
(e)
(f)
0.4 r = 0.69 p = 1.10e-09 0.3
r = 0.36 p = 4.47e-03
0.4
0.2 0.2
0.1
0.4
(c)
0 0.1 0.2 0.3 0.4 0.5
0.2
0
MA knn [nats]
0.6
r = 0.33 p = 9.43e-03
161
(g)
0.1
0.2
0.3
0 0.4
r = 0.76 0.3 p = 9.49e-13
0.2
0.1
0.1 0.1
lin [nats]
0.2
(h)
0.4
0.6
r = 0.66 0.3 p = 5.96e-09
0.2
0
0.1
0.2
0
0
(i)
0.05 0.1 0.15 0.2
0.1 r = 0.57 p = 1.71e-06 0.05
0.1 0.2 0.3 0.4 0.5
lin [nats]
0
0.02
0.06
0.1
lin [nats]
Fig. 2. Scatter plots of pair of linear (lin) and nonlinear (knn) estimates of TE measures computed for each subject in a given experimental condition (REST in light blue, HUT in pink and MA in green). The values of the information transferred from RESP to RR (left column: a, d, g panels), from RESP to SAP (middle column: b, e, h panels), and from SAP to RR (right column: c, f, i) are reported. In each subplot, the solid line epicts the regression line between the two estimators, and the correlation value r as well as its significance level p are reported at the top left of each subplot.
4 Conclusion Our findings confirm that respiratory activity has a reduced influence on cardiovascular activity during stress, owing to the decrease in parasympathetic activity and the increase in the sympathetic one; the latter determines an increase in baroreflex activity during orthostatic stress. The robustness of these results is corroborated by the fact that both model-based and model-free approaches for TE estimation perform similarly in discriminating between stressful and resting conditions. This is further supported by the significant correlation between the TE values obtained using the two approaches. Therefore, our findings support the adoption of the parametric approach for the investigation of physiological control mechanisms in healthy subjects undergoing tasks leading to the activation of the sympathetic system, given its simplicity compared to model-free ones.
162
C. Barà et al.
In this work, a bivariate analysis framework has been used to investigate the interplay between different physiological systems. The impossibility to account for the influence of other systems participating in complex control mechanisms is the main limitation of this approach since does not comprise the use of the more appropriate measure of conditional transfer entropy [6]. Future research should explore the use of simpler and computationally faster model-free estimation methods, e.g., binning or permutationbased approaches [8], and compare their discriminative capability with that of the linear approach also on other databases, including in pathological conditions where nonlinear interaction dynamics are dominant [25]. Acknowledgement. This work was supported by “Sensoristica intelligente, infrastrutture e modelli gestionali per la sicurezza di soggetti fragili” (4FRAILTY) project, funded by Italian MIUR (PON R&I 2014–20, grant no. ARS01 00345) and by SiciliAn MicronanOTecH Research And Innovation CEnter ” SAMOTHRACE” (MUR, PNRR-M4C2, ECS 00000022), spoke 3—Universita‘ degli Studi di Palermo” S2-COMMs-Micro and Nanotechnologies for Smart & Sustainable Communities. R.P. is partially supported by the European Social Fund (ESF) Complementary Operational Programme (POC) 2014/2020 of the Sicily Region.
References 1. Ivanov, P.C.: The new field of network physiology: building the human physiolome. Front. Netw. Physiol. 1, 711778 (2021) 2. Schulz, S., Adochiei, F.-C., Edu, I.-R., Schroeder, R., Costin, H., Bär, K.-J., Voss, A.: Cardiovascular and cardiorespiratory coupling analyses: a review. Philos. Trans. Royal Soc. A: Math. Phys. Eng. Sci. 371(1997), 20120191 (2013) 3. Lizier, J.T.: The Local Information Dynamics of Distributed Computation in Complex Systems. Springer Science & Business Media (2012) 4. Turing, A.M., et al.: On computable numbers, with an application to the entscheidungsproblem. J. Math. 58, 345–363 (1936) 5. Vicente, R., Wibral, M., Lindner, M., Pipa, G.: Transfer entropy—a model-free measure of effective connectivity for the neurosciences. J. Comput. Neurosci. 30, 45–67 (2011) 6. Faes, L., Kugiumtzis, D., Nollo, G., Jurysta, F., Marinazzo, D.: Estimating the decomposition of predictive information in multivariate systems. Phys. Rev. E 91(3), 032904 (2015) 7. Schreiber, T.: Measuring information transfer. Phys. Rev. Lett. 85(2), 461 (2000) 8. Bara, C., Sparacino, L., Pernice, R., Antonacci, Y., Porta, A., Kugiumtzis, D., Faes, L.: Comparison of discretization strategies for the model-free information-theoretic assessment of short-term physiological interactions. Chaos: Interdisc. J. Nonlinear Sci. 33(3), 033127 (2023) 9. Javorka, M., et al.: Towards understanding the complexity of cardiovascular oscillations: Insights from information theory. Comput. Biol. Med. 98, 48–57 (2018) 10. Faes, L., Nollo, G., Porta, A.: Compensated transfer entropy as a tool for reliably estimating information transfer in physiological time series. Entropy 15(1), 198–219 (2013) 11. Porta, A., De Maria, B., Bari, V., Marchi, A., Faes, L.: Are nonlinear model-free conditional entropy approaches for the assessment of cardiac control complexity superior to the linear model-based one? IEEE Trans. Biomed. Eng. 64(6), 1287–1296 (2016) 12. Faes, L., Marinazzo, D., Jurysta, F., Nollo, G.: Linear and non-linear brain– heart and brain– brain interactions during sleep. Physiol. Meas. 36(4), 683 (2015)
Comparison of Linear Model-Based and Nonlinear Model-Free Measures
163
13. Fortrat, J.-O., Yamamoto, Y., Hughson, R.L.: Respiratory influences on non-linear dynamics of heart rate variability in humans. Biol. Cybern. 77(1), 1–10 (1997) 14. Mijatovic, G., Pernice, R., Perinelli, A., Antonacci, Y., Busacca, A., Javorka, M., Ricci, L., Faes, L.: Measuring the rate of information exchange in point-process data with application to cardiovascular variability. Front. Netw. Physiol. 1, 765332 (2022) 15. Kraskov, A., Stögbauer, H., Grassberger, P.: Estimating mutual information. Phys. Rev. E 69(6), 066138 (2004) 16. Schwarz, G.: Estimating the dimension of a model. Ann. Stat. 461–464 (1978) 17. Pernice, R., Javorka, M., Krohova, J., Czippelova, B., Turianikova, Z., Busacca, A., Faes, L., Member, I.: Comparison of short-term heart rate variability indexes evaluated through electrocardiographic and continuous blood pressure monitoring. Med. Biol. Eng. Comput. 57, 1247–1263 (2019) 18. Cooke, W.H., Hoag, J.B., Crossman, A.A., Kuusela, T.A., Tahvanainen, K.U., Eckberg, D.L.: Human responses to upright tilt: a window on central autonomic integration. J. Physiol. 517(2), 617–628 (1999) 19. Sant’anna, I.D., de Sousa, E.B., de Moraes, A.V., Loures, D.L., Mesquita, E.T., da Nobrega, A.C.L.: Cardiac function during mental stress: cholinergic modulation with pyridostigmine in healthy subjects. Clin. Sci. 105(2), 161–165 (2003) 20. Grossman, P., Taylor, E.W.: Toward understanding respiratory sinus arrhythmia: relations to cardiac vagal tone, evolution and biobehavioral functions. Biol. Psychol. 74(2), 263–285 (2007) 21. Triedman, J.K., Saul, J.P.: Blood pressure modulation by central venous pressure and respiration. buffering effects of the heart rate reflexes. Circulation 89(1), 169–179 (1994) 22. Faes, L., Gomez-Extremera, M., Pernice, R., Carpena, P., Nollo, G., Porta, A., BernaolaGalvan, P.: Comparison of methods for the assessment of nonlinearity in short-term heart rate variability under different physiopathological states. Chaos: Interdisc. J. Nonlinear Sci. 29(12), 123114 (2019) 23. Porta, A., Marchi, A., Bari, V., Heusser, K., Tank, J., Jordan, J., Barbic, F., Furlan, R.: Conditional symbolic analysis detects nonlinear influences of respiration on cardiovascular control in humans. Philos. Trans. Royal Soci. A: Math. Phys. Eng. Sci. 373(2034), 20140096 (2015) 24. Porta, A., Guzzetti, S., Borroni, E., Furlan, R., Montano, N., Malliani, A.: Role of the autonomic nervous system in generating non-linear dynamics in short-term heart period variability. Biomedizinische Technik. Biomed. Eng. 51(4), 174–177 (2006) 25. Nollo, G., Faes, L., Antolini, R., Porta, A.: Assessing causality in normal and impaired shortterm cardiovascular regulation via nonlinear prediction methods. Philos. Trans. Royal Soc. A: Math. Phys. Eng. Sci. 367(1892), 1423–1440 (2009)
Kinematic Characterization of Movements During the Tinetti Test Alessandra Raffini1(B) , Francesco Bassi1 , Miloš Ajˇcevi´c1 , Aleksandar Miladinovi´c2 , and Agostino Accardo1 1 Department of Engineering and Architecture, University of Trieste, Via Valerio 10, 34127
Trieste, Italy [email protected] 2 Institute for Maternal and Child Health-IRCCS Burlo Garofolo, Trieste, Italy
Abstract. The improvement in life expectancy has led to a corresponding increase in people suffering from chronic illnesses as well as in subjects at high risk of falling. Various scales exist in literature to evaluate fall risk in ambulatory settings among which the Tinetti Test is the most used. However, only trained healthcare professionals can conduct this test. In order to make this scale available to a growing number of older people outside hospital and to reduce the high inter operator bias in scoring the exercises, a less provider-dependent system is necessary. In this preliminary study, which can be used as a benchmark for future evaluation of individuals at risk of falling, some parameters were extracted from a wireless 3D magnetic-inertial sensor applied on the chest of 30 young healthy participants. Each subject performed four exercises from the Tinetti balance test: arising from a chair (1), standing balance with open (2), and closed eyes (3) and sitting down (4). For exercises (1) and (4), the duration of movement and the maximum angular amplitude were calculated, while for exercises (2) and (3) the fractal dimension and the spectral power were evaluated. The obtained values, directly correlated with the exercises, showed a low variability among subjects, resulting as potential candidates for the characterization of the movement during the Tinetti test, enabling non-expert operators to assess the falling risk. Keywords: Tinetti test · Fall risk assessment · Magneto-Inertial sensor
1 Introduction The advance of medical science and the growing accessibility to innovative treatments have contributed to a remarkable rise in survival rates for numerous diseases resulting in a general increase in life expectancy. However, this improvement in average lifespan has also led to a corresponding increase in the percentage of individuals with chronic illnesses. This fraction of population needs a more sophisticated level of care and presents an elevated risk of falling, defined as a reduced ability to maintain balance while standing or sitting. As a result, significant physical and psychological costs have to be sustained by patients and social security. To address these evolving needs, global healthcare systems © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 164–171, 2024. https://doi.org/10.1007/978-3-031-49062-0_18
Kinematic Characterization of Movements During the Tinetti Test
165
must adopt innovative strategies and establish a framework that prioritizes preventive measures to avoid the worsening of conditions that could have been easily treated in the early stages. The approach to preventive medicine can be implemented through risk assessment via evaluation scales, used in ambulatory settings or hospice, that rely on physical tests. In particular, for assessing fall risk, the literature proposes different scales such as Timed Up and Go (TUG) test (to assess functional mobility and dynamic balance), Berg Balance Scale (BBS) (to assess balance performance), Functional Reach Test (FRT) (to assess dynamic balance performance), and Tinetti Mobility Test (to assess gait and balance performance) [1, 2]. These tools have proven to be flexible enough to evaluate not only healthy elderly population, but also patients with different underlying pathological conditions like, for example, Parkinson’s disease [2], patients receiving hemodialysis [3], and systemic sclerosis [4]. Among all the scales, the Tinetti Mobility Test developed by Tinetti [5], also called Performance Oriented Mobility Assessment (POMA), is the most used. This scale, which can be utilized only by trained healthcare professionals, is composed of a series of scored exercises aimed to assess balance and gait. In the first part, focused on balance (ranged from 0 to 16 points), the patients are asked to sit down, arise from the chair, stand firstly with eyes open then with eyes closed, withstand a nudge on the chest, turn 360°, and sit down. While the second part, targeted on gait (ranged from 0 to 12 points), consists in letting the patient walk and observing step parameters (i.e. step height, step length, continuity symmetry walking stance, amount of trunk sway, path deviation). The sum of the scores obtained in each of the two parts, defines the overall numeric evaluation of fall risk: if below 20 the fall risk is high otherwise the risk is classified as low [6]. In order to make this scale available to a growing number of older people outside hospital and reduce the high inter operator bias in scoring the exercises, it is necessary to implement a system independent of the healthcare provider. The use of specific devices, such as wearable accelerometers/gyroscopes, can be a solution to this problem thanks to the possibility to gather large amount of data related to movements produced during POMA test, which can be processed so as to estimate the Tinetti score [7]. The literature reports other studies about the use of inertial sensors for fall risk assessment, which, however, show great variability in the results mainly due to the placement of the sensors, the features extracted from them and the test used [8]. However, the use of new technologies leads to the need for new standard data to refer to. The aim of this study was to evaluate the possibility of automatic extraction of one or more parameters from a magnetic-inertial sensor that could be directly correlated with the execution of exercises included in the Tinetti test, thus supporting non-expert operators to assess the falling risk. This preliminary study can be used as a benchmark for future evaluation of individuals at risk of falling.
2 Materials and Methods 2.1 Subjects and Tasks A group of 30 healthy young subjects, 15 males and 15 females (aged 21–31, mean 22.7 ± 2.3 years), participated to the study. The subjects, presenting a Tinetti score of 28, performed four different exercises of the Tinetti test: arising from a chair (1), standing
166
A. Raffini et al.
balance with open eyes (2) and with closed eyes (3), and sitting down (4). In order to calibrate the system, before starting the first and the second task, the subject was asked to maintain the resting position for few seconds, sitting with the back leaning against the backrest or standing with arms along sides, respectively. All the enrolled subjects signed the informed consent. 2.2 Acquisition and Analysis For the data acquisition, a wireless magneto-inertial sensor from MTw Awinda Development Kit (Xsens Technology) capable of detecting 3D Euler angles was used. The sensor was placed on the chest of the subject by means of Velcro body strap so that pitch, roll and yaw angles were the rotations around respectively anteroposterior, right-left and top-down directions (Fig. 1).
Fig. 1. Schematic representation of Pitch, Roll and Yaw angles.
A specific software (MT Manager) allows the operator to link the devices to the host and to acquire the data at a sample frequency of 100 Hz. The collected data were processed using a program written in MATLAB® . For the exercises (1) and (4), the evaluated parameters were the duration of the rotation movement and the maximum angular amplitude along the three directions, while for (2) and (3), the parameters evaluated were the fractal dimension, calculated with the Higuchi algorithm [9], and the total spectral power extracted from the Power Spectral Density estimated by using Welch’s method. These values were averaged over all subjects and the coefficient of variation (CV) was evaluated to estimate the inter-subjects variability.
3 Results Figure 2 shows the trends of the median curves of roll, pitch and yaw angles together with twenty-fifth and seventy-fifth percentiles during the execution of exercise (1) and (4). Since the direction of rotation (right/left, clockwise/counterclockwise) is not relevant for the evaluation of the exercise, the angles measured by roll and yaw are referred as their absolute values. The roll and yaw angle graphs have been time-aligned to the pitch angles by setting the instant corresponding to the maximum pitch angle to zero.
Kinematic Characterization of Movements During the Tinetti Test
167
Fig. 2. Trends of median (black), 25th (green) and 75th (red) percentile of absolute values of roll, pitch and yaw for exercise (1), left side, and exercise (4), right side.
As expected, the rotation along the anteroposterior direction (pitch) presents, in both exercises, a very large peak value (till 45° and about 50° for the exercise (1) and (4), respectively) much greater than those in the other directions. Moreover, Table 1 shows a quite low inter-subject variability for both peak and duration values measured by mean of CV in the exercises (1) and (4). Figure 3 shows a typical angle rotation behavior around the three directions during the exercises (2) and (3) recorded in a subject. The amplitude is very small (always under 1°–7°) in all the directions with some fluctuations and values similar in the two situations. Table 2 reports median, 25th percentile, 75th percentile, and relative standard deviation (CV) values of spectral power and fractal dimension calculated in balance exercises with open eyes (exercise 2) and closed eyes (exercise 3). As in the exercises (1) and (4), also in these tests the greatest median spectral power as well as fractal dimension values are showed by pitch, followed by yaw and roll. The coefficients of variation presented a large relative variability for spectral power, especially in the exercise with open eyes. On the contrary, the CV of the fractal dimension of roll, pitch, and yaw is very low (0.05–0.09) in both exercises.
168
A. Raffini et al.
Table 1. Median, 25th percentile, 75th percentile, and relative standard deviation (CV) of Peak and Duration values during exercise (1) and exercise (2). Pitch
Exercise (1)
Exercise (4)
Roll
Yaw
Peak [°]
Duration [s]
Peak [°]
Duration [s]
Peak [°]
Duration [s]
25th percentile
40.4
2.1
2.3
0.1
1.9
0.1
Median
45.0
2.4
8.1
1.1
4.4
0.5
75th Percentile
51.0
2.6
15.7
2.2
11.0
1.2
CV
0.27
0.18
1.16
0.93
1.04
1.14
25th percentile
40.0
2.1
5.0
0.0
28.7
0.89
Median
50.2
2.5
11.5
1.1
16.5
1.1
75th Percentile
58.3
2.7
16.3
1.9
12.7
1.5
CV
0.34
0.25
1.08
1.06
0.48
0.72
Fig. 3. Example of angles of rotation during Open (left side) and Closed Eyes (right side) exer cises recorded in a subject.
Kinematic Characterization of Movements During the Tinetti Test
169
Table 2. Median, 25th and 75th percentiles, Coefficient of Variation (CV) of Spectral Power (SP) and Fractal Dimension (FD) parameters in the balance exercises with open or closed eyes. Pitch
Roll
Yaw
SP [deg2 ] FD [−] SP [deg2 ] FD [−] SP [deg2 ] FD [−] Exercise (2) 25th percentile 0.11 Median
1.08
0.03
1.06
0.04
1.07
0.18
1.14
0.05
1.10
0.12
1.10
75th Percentile 0.35
1.21
0.12
1.15
0.26
1.16
CV
1.42
0.09
2.06
0.05
1.71
0.07
Exercise (3) 25th percentile 0.11
1.07
0.04
1.06
0.04
1.09
Median
0.18
1.11
0.05
1.10
0.09
1.12
75th Percentile 0.29
1.19
0.10
1.14
0.13
1.17
CV
0.08
1.49
0.05
1.32
0.07
0.76
4 Discussion According to the coefficient of variation values in Tables 1 and 2, all the parameters evaluated (peak and duration of rotation for exercise (1) and (4), spectral power and fractal dimension for exercise (2) and (3)) present a low variability and comparable responses among subjects, demonstrating high repeatability of the exercises in young people. The results show that in exercise (1) and (4), the prevailing rotation was, as expected, in the anteroposterior direction (pitch angle) with a peak of 45.0° and a duration of 2.4 s for (1), and 50.2° and 2.5 s for (2), respectively. Roll and yaw presented a lower angular rotation and duration, as showed in Table 1, confirming that both in sitting down and in standing up movements the displacement of the trunk laterally or around the vertical axis is small compared to the anterior posterior one. We would compare these results to those recently reported by Rivolta et al. [7] that used a triaxial accelerometer to register the linear acceleration along vertical, mediolateral and anterior-posterior axes from which amplitude and duration of the movement to execute some Tinetti exercise were evaluated. The sitting down and standing up exercises showed the highest amplitude variation, along the anterior-posterior axis, with a duration, respectively of 2.31 s and 1.47 s for high-risk subjects and of 1.95 s and 1.30 s for low-risk subjects. In our study, the movement duration seems much longer than that of Rivolta et al. even in subjects considered to be at high risk. However, we must consider the fact that in our case the durations are established on the basis of the profiles of the angular rotations which present different amplitudes according to the three rotation axes (Table 1). In the case of Rivolta et al. instead, the durations are calculated from a single acceleration profile that can therefore justify the differences found. A particular note concerns the yaw angle behavior during exercise (4) as depicted in Fig. 2: it is evident that some subjects began the sitting movement turning around its vertical axis before the operator gave the command to sit, so influencing the initial
170
A. Raffini et al.
trait of the signal. However, the amplitude of the movement was correctly estimated discarding the first part of the movement. The results related to spectral power and fractal dimension evaluated for exercise (2) and (3) show that spectral power of pitch and roll is the same for both exercises (0.18 pitch, 0.05 roll), while for yaw there is a slight difference (0.12 open eyes, 0.09 closed eyes), presenting a higher value in the exercise with open eyes. The fractal dimension presents, for all the angles in both conditions, a median value slightly greater than 1 (between 1.10 and 1.14). Since fractal dimension is a measure of the complexity of a time series with value between 1 (smooth fluctuations like those recorded in our subjects) and 2 (abrupt fluctuations), it could be compared with sample entropy (another parameter utilized to measure complexity) that was used in [7]. In the latter paper, during standing balance tasks the sample entropy was normalized to the theoretical sample entropy value of a white Gaussian noise with the same variance and the values obtained for open eyes and closed eyes task were of 0.975 (high risk subjects) and 0.992 (low risk subjects), 0.985 (high risk subjects) and 0.994 (low risk subjects), respectively. Since greater values can be associated with greater signal complexity, we expect that in subjects with a greater risk of falling, the value of the fractal dimension will increase as well as entropy increased.
5 Conclusions The aim of this study was to evaluate the possibility of automatically extracting one or more parameters from a magnetic-inertial sensor that could be directly correlated with the quality of execution of the Tinetti exercises (i.e. with the Tinetti score), enabling nonexpert operators to utilize this test in practice. The results show that some parameters can well characterize the movement during the exercises in a population of young healthy subjects. In particular the peak and the duration of movements in the anterior posterior direction for exercise (1) and exercise (4) and the spectral power and the fractal dimension for exercise (2) and (3) seems to be good candidates showing a low variability among subjects. Moreover, since our system is based on a 3D magneto-inertial sensor, it allows to detect in precise and accurate way the angular rotations distinguishing the movement components along the three orthogonal directions. However, this is a preliminary study that can be used in the future as a benchmark for evaluations of individuals at risk of falling. The next step will concern the application of the same protocol to elderly people.
References 1. Strini, V., Schiavolin, R., Prendin, A.: Fall risk assessment scales: a systematic literature review. Nurs. Rep. 11, 430–443 (2021). https://doi.org/10.3390/nursrep11020041 2. Sebastia-Amat, S., Tortosa-Martínez, J., Pueo, B.: The use of the static posturography to assess balance performance in a Parkinson’s disease population. Int. J. Environ. Res. Public Health 20, 981 (2023). https://doi.org/10.3390/ijerph20020981 3. Zanotto, T., et al.: Association of postural balance and falls in adult patients receiving haemodialysis: a prospective cohort study. Gait Posture 82, 110–117 (2020). https://doi.org/ 10.1016/j.gaitpost.2020.08.128
Kinematic Characterization of Movements During the Tinetti Test
171
4. Yakut H, Özalevl˙i S, B˙iRl˙iK AM (2021) Postural balance and fall risk in patients with systemic sclerosis: A cross-sectional study. Arch Rheumatol 36:167–175. https://doi.org/10.46497/Arc hRheumatol.2021.8259 5. Tinetti, M.E.: Performance-oriented assessment of mobility problems in elderly patients. J. Am. Geriatr. Soc. 34, 119–126 (1986). https://doi.org/10.1111/j.1532-5415.1986.tb05480.x 6. Tinetti, M.E., Franklin Williams, T., Mayewski, R.: Fall risk index for elderly patients based on number of chronic disabilities. Am. J. Med. 80, 429–434 (1986). https://doi.org/10.1016/ 0002-9343(86)90717-5 7. Rivolta, M.W., et al.: Evaluation of the Tinetti score and fall risk assessment via accelerometrybased movement analysis. Artif. Intell. Med. 95, 38–47 (2019). https://doi.org/10.1016/j.art med.2018.08.005 8. Montesinos, L., Castaldo, R., Pecchia, L.: Wearable inertial sensors for fall risk assessment and prediction in older adults: a systematic review and meta-analysis. IEEE Trans. Neural Syst. Rehabil. Eng. 26, 573–582 (2018). https://doi.org/10.1109/TNSRE.2017.2771383 9. Higuchi, T.: Approach to an irregular time series on the basis of the fractal theory. Physica D 31, 277–283 (1988). https://doi.org/10.1016/0167-2789(88)90081-4
A TMS-EEG Pre-processing Parameters Tuning Study Elena Bondi1,2(B) , Viviana Pescuma2 , Yara Massalha3 , Marta Pizzolante4 , Alice Chirico4 , Giandomenico Schiena2 , Anna Maria Bianchi1(B) , Paolo Brambilla2,3 , and Eleonora Maggioni1 1 Department of Electronics Information and Bioengineering, Politecnico Di Milano, Milan,
Italy {elena.bondi,annamaria.bianchi}@polimi.it 2 Department of Neurosciences and Mental Health, Fondazione IRCCS Ca’ Granda Ospedale Maggiore Policlinico, Milan, Italy 3 Department of Pathophysiology and Transplantation, University of Milan, Milan, Italy 4 Department of Psychology, Università Cattolica del Sacro Cuore, Milan, Italy
Abstract. The integration of transcranial magnetic stimulation and electroencephalography (TMS-EEG) has been sought to explore connectivity and cortical excitability in healthy and pathological subjects. Although the number of studies in this field is continuously increasing, there is still no consensus on the pre-processing pipeline. The goal of this work is to study the effect of two preprocessing parameters, (i) the interpolation window size and (ii) the regularization parameter of the source-estimate-utilizing noise-discarding (SOUND) algorithm, on the resulting TMS-evoked potentials (TEPs) and five peaks of interest. The comparison of four combinations of parameters showed the effect of the parameters on the TMS residual artifacts, which resulted in major differences between TEPs in the early windows after the TMS pulses. The work showed how the interpolation window size and the regularization parameter influence the TEPs results and proposed a combination of parameters that retrieve information of interest, decreasing the amount of noise in data. Keywords: TMS-EEG · Cortical excitability · Parameters tuning · Pre-processing
1 Introduction Electroencephalography (EEG) is a non-invasive technique that has been effectively coupled with transcranial magnetic stimulation (TMS) since 1996 [1] to investigate the cortical excitability and effective connectivity in normal [2, 3] and pathological brains [4, 5]. TMS-EEG has been mostly used to study motor cortex stimulation [6, 7], but also other cortical areas such as the dorsolateral prefrontal cortex (DLPFC) [4, 8]. A singlepulse TMS induces a depolarization over the underlying neurons that can be assessed as P. Brambilla and E. Maggioni—These authors contributed equally to this work. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 172–179, 2024. https://doi.org/10.1007/978-3-031-49062-0_19
A TMS-EEG Pre-processing Parameters Tuning Study
173
a series of positive and negative EEG deflections called TMS-evoked potentials (TEPs), which are sensitive to stimulation location [9] and stimulation parameters [10]. Although interesting information can be derived from the integration of TMS and EEG, from a methodological point of view the combination of the two techniques shows some challenges. The artifacts present in the TMS-EEG signal have been characterized by different studies and different online and offline approaches have been proposed to increase the signal-to-noise ratio [9, 11–13]. However, there is no gold standard on the acquisition or pre-processing pipelines, even though there is a consensus in the community on the promotion of standardization [14]. Up to now, different methodologies have been proposed to reduce the artifacts introduced by the TMS [15–18] and few studies have just started comparing different pipelines [19, 20]. A common consensus on the pre-processing pipeline has been reached for the interpolation of the signal in the few milliseconds after the TMS pulse to eliminate the TMS-pulse artifact. Also, an increasing number of studies are using the source-estimate-utilizing noise-discarding (SOUND) algorithm [18], which seems very promising in eliminating TMS-EEG artifacts [4, 19, 21]. The goal of this study is to quantitatively assess the impact of two pre-processing parameters, (i) the interpolation window size (w) and (ii) the regularization parameter of the SOUND algorithm (λ), on the TEPs obtained by 6 healthy controls (HCs) that underwent a resting-state TMS-EEG session where the left DLPFC was stimulated with single-pulse TMS stimuli.
2 Methods 2.1 Participants The dataset was collected from 6 HCs (3 females and 3 males, mean age 28.33 ± 4.13 years). Exclusion criteria for all participants comprised mental retardation (intelligence quotient 0.80). Secondly, generalized linear models (GLMs, p < 0.05) were used to evaluate the effect of the pre-processing parameters (w and λ), using the subject factor as a fixed effect, on (i) the variance of the ICs discarded in the first round of ICA, (ii) the percentage of ICs discarded during the first round of ICA, (iii) the amplitude, and (iv) the latency of the peaks of interest extracted from the TEPs at electrodes of interest. Moreover, the variability among the processed data resulting from the four parameters sets was investigated with an absolute measure of deviation, the quartile deviation, of the amplitude and latency of peaks extracted. The values were inserted into a Friedman test (p < 0.05) to evaluate statistical differences among the four parameters sets.
3 Results In the visual inspection of the GA-TEPs obtained using the four parameters sets (M1 (w = 10, λ = 0.01), M2 (w = 10, λ = 0.1), M3 (w = 5, λ = 0.01), and M4 (w = 5, λ = 0.1)), five peaks were visible at E12 and E65 (Fig. 2). Specifically, P25, N40, P60, N100, and P180 were found at E12, while P40, N60, P90, N110, and P180 were found at E65. The temporal correlation results in the first time window 0–20 ms (Fig. 3) showed a moderate correlation between combinations of parameters with the same w, and slight
176
E. Bondi et al.
to virtually no correlation for the other comparisons. In the other time windows, the coefficient values increased, usually being moderate or substantial. From the GLM analyses, a significant effect of the window length (w) was found for the percentage of variance discarded by the first ICA round (p = 0.032, t = −2.35), the percentage of ICs discarded during the same round (p < 0.001, t = −4.36), and the latency of P180 found at E65 (p = 0.016, t = 2.69). A significant effect of λ was found for the amplitude of N100 at E12 (p = 0.011, t = 2.89). Although the Friedman test did not report a significant result, the variability analyses performed on the amplitude and latency of peaks of interest reported lower values of averaged quartile deviation for M2 and M4 (4.08 and 4.06, respectively) than for M1 and M3 (6.11 and 6.40, respectively).
Fig. 2. Grand-average TMS-evoked potentials (GA-TEP) of four parameters sets obtained at E12 (a) and E65 (b).
Fig. 3. Temporal correlation values (ρ) calculated for all pairs of pipelines, between all pairs of channels’ time series in different time windows.
A TMS-EEG Pre-processing Parameters Tuning Study
177
4 Discussion Despite the increasing body of work around the TMS-EEG data, both from methodological and clinical points of view, there is still no gold standard procedure for pre-processing TMS-EEG data. The main reason is the difference in terms of equipment, and therefore in the signal-to-noise ratio of data. For this study, we proposed a pipeline similar to others already reported [4, 19, 21] which involves the use of the SOUND and SSP-SIR algorithms. We decided to focus on two steps of the analysis, the interpolation around the TMS pulses to eliminate the large TMS-pulse artifacts and the regularization parameter λ used in the SOUND algorithm. The interpolation step is the most common step among the pipelines proposed in the literature. However, there is no consensus on the length of the window, due to the differences in terms of artifacts induced by the TMS in different studies [12, 13]. The regularization parameter, which is inversely proportional to the signal-to-noise ratio, is chosen according to data quality. For this work, we tested two values of λ, used in previous studies [4, 17], and study their effect on data. The results of all the pre-processing parameters sets, in terms of TEPs, were in line with earlier findings [4, 9, 25–27], showing that the pipeline proposed can eliminate at least in part - the artifacts affecting the signal, increasing the signal-to-noise ratio. The choice of parameters affected the resulting TEP in the first 20 ms more than in other intervals, as shown by the correlation values. Specifically, w seemed to influence the TEPs more than λ, while in the late windows results were comparable. This finding can also be supported by a visual inspection of the GA-TEPs, where methods with smaller windows reported similar results, but higher amplitudes with respect to the other methods in the first 20 ms. A large interpolation window can also affect the first round of ICA, reducing the number of TMS-pulse artifactual components that the user has to discard and also the percentage of variance, suggesting that a large window can eliminate the TMS-induced artifact without cutting out the early peaks. The parameter λ influences the repeatability of the results, showing that a higher λ results in a lower variability of TEPs peaks. Overall, these findings suggest that the parameters influence the results, particularly in the first 20 ms, the most affected by TMS-induced artifacts. Also, a larger window could be preferable to diminish the TMS-pulse artifact, reducing the probability of a user’s error in the manual selection of the ICs. Regarding the parameter λ = 0.1, it showed a low inter-subject variability, showing a better reproducibility of the results. The main limitations of this study are the small dataset, which could be enriched in the future, and the low sampling frequency of the EEG amplifier, which could have increased the TMS-pulse artifacts. However, our qualitative and quantitative comparisons of the TEPs support the ability of the proposed pre-processing pipeline - regardless of the parameters - in retrieving the information of interest according to previous results.
5 Conclusions The purpose of the current study was to compare the results of four different sets of pre-processing parameters on TMS-EEG data, specifically evaluating the effect of two parameters, the interpolation window and the regularization parameter of the SOUND algorithm, on EEG data collected from a pilot HC sample during single-pulse TMS
178
E. Bondi et al.
stimuli targeting the left DLPFC. The research has shown that the interpolation window and the regularization parameter influence the residual TMS-related artifactual components and the results of the TEP analysis, in particular in the early window. The results suggested that an interpolation window of 10 ms and a regularization parameter of 0.1 reduce the residual TMS-pulse artifacts and the inter-variability of peaks of TEPs among subjects. The insights gained from this study may support further studies in the choice of parameters and lead to a definition of a standard pre-processing pipeline. Acknowledgements. This work was supported by the BIAL Foundation (SUBRAIN project “The origin of the sublime power in the brain: an integrated EEG-TMS study”, Grant no. 288/20 to EM).
References 1. Ilmoniemi, R.J., Virtanen, J., Ruohonen, J., Karhu, J., Aronen, H.J., Näätänen, R., Katila, T.: Neuronal responses to magnetic stimulation reveal cortical reactivity and connectivity. Neuroreport 8(16), 3537–3540 (1997) 2. Thut, G., Pascual-Leone, A.: A review of combined TMS-EEG studies to characterize lasting effects of repetitive TMS and assess their usefulness in cognitive and clinical neuroscience. Brain Topogr. 22, 219–232 (2010) 3. Rogasch, N.C., Fitzgerald, P.B.: Assessing cortical network properties using TMS–EEG. Hum. Brain Mapp. 34(7), 1652–1669 (2013) 4. Bagattini, C., et al.: Predicting alzheimer’s disease severity by means of TMS–EEG coregistration. Neurobiol. Aging 80, 38–45 (2019) 5. Massimini, M., Ferrarelli, F., Sarasso, S., Tononi, G.: Cortical mechanisms of loss of consciousness: insight from TMS/EEG studies. Arch. Ital. Biol. 150(2/3), 44–55 (2012) 6. Paus, T., Sipila, P., Strafella, A.: Synchronization of neuronal activity in the human primary motor cortex by transcranial magnetic stimulation: an EEG study. J. Neurophysiol. 86(4), 1983–1990 (2001) 7. Veniero, D., Brignani, D., Thut, G., Miniussi, C.: Alpha-generation as basic responsesignature to transcranial magnetic stimulation (TMS) targeting the human resting motor cortex: a TMS/EEG co-registration study. Psychophysiology 48(10), 1381–1389 (2011) 8. Kähkonen, S., Komssi, S., Wilenius, J., Ilmoniemi, R.J.: Prefrontal transcranial magnetic stimulation produces intensity-dependent EEG responses in humans. Neu roimage 24(4), 955–960 (2005) 9. Tremblay, S., Rogasch, N.C., Premoli, I., Blumberger, D.M., Casarotto, S., Chen, R., Di Lazzaro, V., Farzan, F., Ferrarelli, F., Fitzgerald, P.B., et al.: Clinical utility and prospective of TMS–EEG. Clin. Neurophysiol. 130(5), 802–844 (2019) 10. Casarotto, S., Romero Lauro, L.J., Bellina, V., Casali, A.G., Rosanova, M., Pigorini, A., Defendi, S., Mariotti, M., Massimini, M.: EEG responses to TMS are sensitive to changes in the perturbation parameters and repeatable over time. PloS One 5(4), e10,281 (2010) 11. Ilmoniemi, R.J., Kici´c, D.: Methodology for combined TMS and EEG. Brain Topogr. 22, 233–248 (2010) 12. Rogasch, N.C., et al.: Analysing concurrent transcranial magnetic stimulation and electroencephalographic data: A review and introduction to the open-source Tesa software. Neuroimage 147, 934–951 (2017) 13. Varone, G., et al.: Real-time artifacts reduction during TMS-EEG co-registration: a comprehensive review on technologies and procedures. Sensors 21(2), 637 (2021)
A TMS-EEG Pre-processing Parameters Tuning Study
179
14. Hernandez-Pavon, J.C., Veniero, D., Bergmann, T.O., Belardinelli, P., Bortoletto, M., Casarotto, S., Casula, E.P., Farzan, F., Fecchio, M., Julkunen, P., et al.: Tms combined with EEG: recommendations and open issues for data collection and analysis. Brain Stimul. (2023) 15. Wu, W., et al.: Artist: a fully automated artifact rejection algorithm for single-pulse TMS-EEG data. Hum. Brain Mapp. 39(4), 1607–1625 (2018) 16. Mutanen, T.P., Kukkonen, M., Nieminen, J.O., Stenroos, M., Sarvas, J., Ilmoniemi, R.J.: Recovering TMS-evoked EEG responses masked by muscle artifacts. Neuroimage 139, 157– 166 (2016) 17. Mutanen, T.P., Metsomaa, J., Liljander, S., Ilmoniemi, R.J.: Automatic and robust noise suppression in EEG and MEG: the sound algorithm. Neuroimage 166, 135–151 (2018) 18. Mutanen, T.P., Biabani, M., Sarvas, J., Ilmoniemi, R.J., Rogasch, N.C.: Source-based artifactrejection techniques available in Tesa, an open-source TMS–EEG tool-box. Brain Stimul.: Basic Transl. Clin. Res. Neuromodul. 13(5), 1349–1351 (2020) 19. Bertazzoli, G., Esposito, R., Mutanen, T.P., Ferrari, C., Ilmoniemi, R.J., Miniussi, C., Bortoletto, M.: The impact of artifact removal approaches on TMS–EEG signal. NeuroImage 239, 118, 272 (2021) 20. Rogasch, N.C., Biabani, M., Mutanen, T.P.: Designing and comparing cleaning pipelines for TMS-EEG data: a theoretical overview and practical example. J. Neurosci. Methods 109494 (2022) 21. Bortoletto, M., Bonzano, L., Zazio, A., Ferrari, C., Pedull‘a, L., Gasparotti, R., Miniussi, C., Bove, M.: Asymmetric transcallosal conduction delay leads to finer bimanual coordination. Brain Stimul. 14(2), 379–388 (2021) 22. Sekiguchi, H., Takeuchi, S., Kadota, H., Kohno, Y., Nakajima, Y.: TMS-induced artifacts on EEG can be reduced by rearrangement of the electrode’s lead wire before recording. Clin. Neurophysiol. 122(5), 984–990 (2011) 23. Delorme, A., Makeig, S.: Eeglab: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J. Neurosci. Methods 134(1), 9–21 (2004) 24. Shrout, P.E.: Measurement reliability and agreement in psychiatry. Stat. Methods Med. Res. 7(3), 301–317 (1998) 25. Lioumis, P., Kici´c, D., Savolainen, P., Mäkelä, J.P., Kähkänen, S.: Reproducibility of TMS— evoked EEG responses. Human Brain Map. 30(4), 1387–1396 (2009) 26. Farzan, F., Bortoletto, M.: Identification and verification of a ‘true’ TMS evoked potential in TMS-EEG. J. Neurosci. Methods 109651 (2022) 27. Biermann, L., Wunram, H.L., Pokorny, L., Breitinger, E., Großheinrich, N., Jarczok, T.A., Bender, S.: Changes in the TMS-evoked potential n100 in the dorsolateral prefrontal cortex as a function of depression severity in adolescents. J. Neural Trans. 129(11), 1339–1352 (2022)
Non-contact Biopotential Amplifier with Capacitive Driven Right Leg Circuit Dino Cindri´c(B) , Luka Klai´c, Antonio Staneši´c, and Mario Cifrek Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia {dino.cindric,luka.klaic}@fer.hr
Abstract. In this paper, the development and measurement results of active capacitive electrodes are presented. A brief overview and implementation of a capacitive driven right leg circuit (CDRL) for biopotential measurements are given. The obtained results showed the feasibility of the developed system for capacitive measurement of ECG signal. When CDRL is used, the common-mode rejection was reduced up to 11.3 dB compared to the case with no CDLR. Finally, the results of filtering in the digital domain on the acquired signals are presented in time and frequency domains. Keywords: Biopotential measurements · Capacitive electrodes · Capacitive driven right leg · Analog frontend · Digital filtering
1 Introduction and Related Work Non-contact measurements of electrophysiological signals are increasingly becoming more attractive in the domain of wearable consumer electronic devices, as well as in clinical applications. Interesting possible applications in the clinical environment are non-invasive monitoring of the health status of newborns and patients in intensive care, where the placement of classical, contact electrodes can cause a whole range of problems such as skin irritation, the possibility of injury, and impracticality and necessary preparation before electrode placement [1]. In the automotive industry, the possibility of using non-contact electrodes for measuring the driver’s electrophysiological parameters during driving is also being researched [2], which, in synergy with modern methods of digital signal processing and machine learning, can play a significant role in increasing the safety not only of drivers but of all traffic participants. The non-contact capacitive measurements come with their own set of problems and engineering challenges, mainly ensuring the signal integrity, signal-to-noise ratio, common-mode rejection ratio, high input impedance and measurements repeatability, and all of that in the small form factor in order to minimize the effect of the measuring system on the human body. Many of these problems are not completely solved and further research on capacitive measurements of electrophysiological signals shows great potential. In measurements where the standard, contact electrodes are used, the driven right leg (DRL) circuit has been used for a long time for the purpose of common mode © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 180–188, 2024. https://doi.org/10.1007/978-3-031-49062-0_20
Non-contact Biopotential Amplifier with Capacitive Driven Right Leg Circuit
181
noise attenuation [3]. The principle of operation of the DRL circuit is based on the inversion of the common mode noise present on the measurement electrodes and the transmission of the processed signal back into the body through a separate electrode. To estimate the amount of common mode noise on the measurement electrodes, the output voltage is first amplified by a preamplifier, typically with low gain, and then passed through a resistor to the input of an operational amplifier connected in an inverting summing configuration, before being amplified by an instrumentation or differential amplifier. The typical gain of the DRL circuit is in the range of 50–1000, and to limit the maximum amount of current that can be returned to the body, a resistor of typical value 10–100 k is connected in series between the output of the amplifier and the electrode. In [4], the DRL used in contact measurement of biopotentials has been adapted for use in a capacitive implementation by returning the inverted and amplified signal back into the body capacitively via conductive fabric that is not in direct contact with the body. The conducted experiment involved comparing a passive reference electrode directly connected to the reference point (ground) of the system to a DRL with gains of 10, 100, and 1000. The obtained results show a reduction in common-mode powerline interference by 40 dB compared to the system without DRL or reference electrode, with a DRL gain of 1000. Even when the DRL gain was set to 10, a significant improvement of 20–30 dB is shown compared to the passive reference electrode. Under certain conditions, high DRL gain values can lead to instability and oscillations in the operation of capacitive DRL systems. In [5], the primary causes of potential CDRL system instability are identified as variable tissue impedance, the amount of series resistance of the protective resistor, characteristics of non-ideal operational amplifiers (poles at higher frequencies, the frequency response in a unity gain amplifier configuration, maximum capacitive load, etc.) and parasitic capacitances. Similar implementations of CDRL systems based on classical DRL used with contact electrodes have been realized in [6–8]. The thickness of the subject’s clothing and the size of the effective surface area of the electrodes are emphasized as the greatest sources of signal quality variability. The CDRL implemented in [7] is identical to that in [6], and the dependence of signal quality on the signal-to-noise ratio for different values of cloth thickness is also reported. The OPA124 amplifier with a gain of 1000 was used to implement the system in [7], and the INA114 instrumentation amplifier with a gain of 5000 was used in [8]. In all cases, it was experimentally demonstrated that using topologically very simple CDRL systems provides an advantage compared to systems without an additional electrode, as well as to those with a passive reference electrode. In [9], three CDRL electrodes with different sensing surfaces were fabricated. For the smallest electrode with an area of 1650 cm2 , the measured level of common-mode interference was 70.96 dB, for the electrode with an area of 3300 cm2 it was 59.13 dB, and for the largest electrode with an area of 4950 cm2 , the interference level was 43.44 dB higher than the useful signal. Compared to a passive electrode connected to the system’s reference point, the CDRL electrode achieved 46.30 dB higher common-mode interference attenuation. In [10], one of the main conclusions of the work is that the addition of the CDRL electrode, even with the large surface area (960 cm2 ) used in the experiment, cannot
182
D. Cindri´c et al.
completely eliminate the influence of the powerline interference, and additional filtering is required. Motion artifacts are highlighted as a significant problem because their occurrence in some cases can cause a preamplifier saturation.
2 Electronics System Design The design of the electronic system of the active capacitive electrode is based on the previous work done in [11], with improvements presented here. For the active electrode preamplifier, Texas Instrument’s LMP7721 precision amplifier is used. Figure 1. Shows the designed preamplifier alongside the passive network. The key difference in the preamplifier design compared to previous work in [11] is the addition of a buffer amplifier between the preamplifier output and guard ring which encompasses the preamplifier’s input. In this configuration, leakage currents are closing through the low-impedance guard ring rather than through the input passive network. The electric potential of the guard ring is determined by the resistor R5 , which is in this case set to 100 . Additionally, implemented preamplifier supports both transimpedance and bootstrap configuration [11].
Fig. 1. Preamplifier circuit
After the preamplifier, the next component of the active electrode electronic signal chain is a passband filter, which is implemented as a cascade of 2nd-order high-pass and 4th-order low-pass filters. The cut-off frequency of the high-pass filter is set at 0.5 Hz with a gain of 20 dB, while the cut-off frequency of the low-pass filter is set to 150 Hz with a gain of 40 dB. Both filters are based on the standard Sallen-Key topology and the values of passive components that determine the cut-off frequencies are identical as in [11].
Non-contact Biopotential Amplifier with Capacitive Driven Right Leg Circuit
183
The final component of analog frontend is the bandstop filter implemented in the twinT configuration with the central frequency set to 50 Hz. Additionally, two operational amplifiers LMV772MA were used as buffer amplifiers between the passive network at the output of the low-pass filter and the electrode’s signal output. The amplitude-frequency plot of the designed active electrode is shown in Fig. 2.
Fig. 2. Amplitude-frequency characteristic of the active capacitive electrode
The 3D model of the printed circuit board (PCB) of the active electrode is shown in Fig. 3. The PCB is implemented in 4 layers in the FR-4 technology – the top layer was used for analog signal routing, two inner layers as a reference ground plane and power plane (2.5 V) respectively, and the bottom layer as a sensing surface of the electrode. In the guard ring (shield) area around the input terminals of the preamplifier (U1), the solder mask was removed to further reduce the influence of parasitic capacitance and leakage currents. The signal from the sensing part of the electrode, which is located on the bottom layer, is fed to the amplifier input through via. The designed capacitive DRL system is simple and it consists of a cascade of buffer and inverting amplifier with a gain set to 10 (20 dB). The two-layer circular CDRL PCB with a radius of 40 mm is shown in Fig. 4. On the bottom layer, the solder mask was removed, so this electrode can be used as DRL in both contact and non-contact measurement systems.
184
D. Cindri´c et al.
Fig. 3. Active electrode printed circuit board
Fig. 4. CDRL printed circuit board
3 Results For the data acquisition from two capacitive electrodes, the USB-6211 data acquisition system and LabView graphical programming environment from National Instruments were used. Output signals from capacitive electrodes were connected to analog inputs of the USB-6211 and differentially amplified with the gain set to 1 (0 dB). For the data analysis and signal processing, MATLAB was used. The sampling frequency is set to 1 kHz. Electrodes were used in the bootstrap configuration [11]. The measurement system block diagram is shown in Fig. 5. Figure 6 shows one of the retrieved measurements of the ECG signal in the time and frequency domain.
Non-contact Biopotential Amplifier with Capacitive Driven Right Leg Circuit
185
Fig. 5. Measurement system block diagram
Fig. 6. Time and frequency domain plot of the acquired raw ECG signal
The y-axis on the time domain plot shows the measured differential voltage from the two electrodes, while the y-axis on the frequency domain plot shows the power spectrum value expressed in decibels (dB) as a result of performing the Fourier transform. In t = 10 s, fingers (thumbs) were placed separately on each of the electrodes, and in t = 36 s they were removed from the electrodes. Even though the fingers were physically in contact with the electrode’s PCB, there was no direct electrical contact with the sensing surface because of the solder mask existence. The presence of a 50 Hz powerline interference and its higher harmonics can clearly be observed in the signal spectrum. Figure 7 shows the time and frequency domain plot of the signal when the digital filters with the following specifications were applied: • high-pass filter with a cut-off frequency of 0.05 Hz, stopband attenuation of 60 dB and steepness factor of 0.85 (value of 1 represents the ideal brick wall filter), • low-pass filter with a cut-off frequency of 150 Hz, stopband attenuation of 60 dB and steepness factor of 0.85, and finally,
186
D. Cindri´c et al.
• three bandstop filters with central frequencies at fundamental and first two higher harmonics of the powerline interference, i.e., at 50 Hz, 100 Hz and 150 Hz. The difference between the upper and lower passband frequencies for all bandstop filters was set to 2 Hz with a stopband attenuation of 80 dB.
Fig. 7. Filtered ECG signal
Figure 8 shows the output differential voltage and its frequency spectrum from two electrodes when ECG is not measured. The maximum observed peak-to-peak noise at the output of the system is in the range of 70 mV.
Fig. 8. Noise at the output when ECG is not measured
Non-contact Biopotential Amplifier with Capacitive Driven Right Leg Circuit
187
Finally, Fig. 9 shows the frequency spectrum comparison of the acquired ECG signal with (blue color) and without (red color) the presence of CDRL electrode. In this example, it can be observed that CDRL attenuates the fundamental component, but also all of the higher harmonics of the powerline interference – the power of the 50 Hz component is 4.3 dB lower, and the power of interference at 100 Hz is by 11.3 dB lower when the CDRL is used compared to a system which has no CDRL.
Fig. 9. Comparison of the acquired signal spectrum with (blue) and without (red) CDRL
4 Conclusion and Further Work In this paper, the feasibility of non-contact, capacitive ECG measurements with the presented electronic system was demonstrated. Although significant attention was given to the design of the analog circuitry of the active measurement electrode, as well as the CDRL electrode, in order to reduce the unwanted external influences, the measurements showed that digital signal processing is also necessary to obtain diagnostically significant data, one of the main problems being the repeatability of measurements. The measured results show the positive influence of the CDRL electrode, and further development of the system can focus on simplifying the electronic system of the active measurement electrode, as well as developing algorithms for digital signal processing and machine learning for diagnostic information extraction from capacitively obtained measurements. There is also a significant potential for developing a dedicated low-power wearable hardware platform for data acquisition and transfer of capacitive measurements, as well as its software support to make the system more reliable and easier to use for the end-user.
References 1. Atallah, L., Serteyn, A., Meftah, M., Schellekens, M., Vullings, R., Bergmans, J.W., Osagiator, A., Oetomo, S.B.: Unobtrusive ECG monitoring in the NICU using a capacitive sensing array. Physiol. Meas. 35(5), 895–913 (2014). https://doi.org/10.1088/0967-3334/35/5/895 2. Uguz, D.U., Dettori, R., Napp, A., Walter, M., Marx, N., Leonhardt, S., Hoog Antink, C.: Car seats with capacitive ECG electrodes can detect cardiac pacemaker spikes. Sensors 20(21), 6288 (2020) 3. Winter, B.B., Webster, J.G.: Driven-right-leg circuit design. IEEE Trans. Biomed. Eng. BME 30(1), 62–66 (1983)
188
D. Cindri´c et al.
4. Lim, Y.G., Gih, S.C.: Capacitive driven-right-leg grounding in Indirect-contact ECG measurement. In: Annual International Conference of the IEEE Enginerring in Medicine and Biology Society, pp 1250–1253 (2010) 5. Haberman, M.A., Spinelli, E.M., García, P.A., Guerrero, F.N.: Capacitive driven-right-leg circuit design. Int. J. Biomed. Eng. Technol. 17(2), 115–126 (2015). https://doi.org/10.1504/ IJBET.2015.068051 6. Baek, H.J., Chung, G.S., Kim, K.K., Park, K.S.: A smart health monitoring chair for nonintrusive measurement of biological signals. IEEE Trans. Inf. Technol. Biomed. 16(1), 150–158 (2012). https://doi.org/10.1109/TITB.2011.2175742 7. Rachim, V.P., Chung, W.-Y.: Wearable noncontact armband for mobile ECG monitoring system. In: IEEE Trans. Biomed. Circ. Syst. 10(6), 1112–1118 (2016) 8. Kim, K.K., Lim, Y.K., Park, K.S.: Common mode noise cancellation for electrically noncontact ECG measurement system on a chair. In: IEEE Engineering in Medicine and Biology 27th Annual Conference, Shanghai, China, pp. 5881–5883 (2005) 9. Bednar, T., Babusiak, B., Labuda, M., Smetana, M., Borik, S.: Common-mode voltage reduction in capacitive sensing of biosignal using capacitive grounding and DRL electrode. Sensors 21(7), 2568 (2021). https://doi.org/10.3390/s21072568 10. Babusiak, B., Stefan, B., Maros, S., Ladislav, J.: Smart sheet design for electrocardiogram measurement. Inform. Technol. Biomed. 507–517 (2019) 11. Cindri´c, D., Staneši´c, A., Cifrek, M.: Analog frontend for ECG/EMG capacitive electrodes. In: 44th International Convention on Information, Communication and Electronic Technology (MIPRO), pp. 363–366 (2021). https://doi.org/10.23919/MIPRO52101.2021.9597199
Relationship Between Personality and Kinematic Parameters of Handwriting Francesco Bassi1(B) , Alessandra Raffini1 , Miloš Ajˇcevi´c1 , Aleksandar Miladinovi´c2 , Lisa Di Blas3 , and Agostino Accardo1 1 Department of Engineering and Architecture, University of Trieste, Via Valerio 10, 34127
Trieste, Italy [email protected] 2 Institute for Maternal and Child Health-IRCCS Burlo Garofolo, Trieste, Italy 3 Department of Life Science, University of Trieste, Via Weiss 2, 34128 Trieste, Italy
Abstract. Motor and cognitive systems are largely involved in producing handwriting that develops with age becoming more and more personalized until it reaches a style proper to the subject. This fact has led graphologists to assert that by examining the handwriting it is possible to somehow trace the personality of the writer. Many studies have been carried out to demonstrate this assumption but they are all based on the graphic examination of the tract left on the sheet of paper. On the other hand, recently it has been possible to examine writing through the use of digital tablets capable of providing information also on the kinematic of the movement, extracting parameters used to examine in particular dysgraphia and some neurological pathologies. Aim of this study was to determine possible relationships between kinematic parameters extracted using digital tablets and personality traits. Sixty-one subjects took part in the study, executing three writing tasks (fast and accurate writing of an Italian phrase and fast sequence of cursive lowercase letters “lelele” without pen lifting for 30 s) and a personality test (IPIP-NEO-120). The linear regression between each of fourteen characteristic of handwriting and each of the five personality traits was computed. The results showed that four out of five main psychological tracts presented a linear relation with one or more kinematic characteristics. This study offers a first glance at a complex series of correlations, which will be investigated in future researches. Keywords: Handwriting · Big Five · Kinematic parameters
1 Introduction Handwriting is a complex process characterized by the involvement of cognitive and motor systems and, at the same time, it is a way to express one’s personality. In order to describe individual differences in personality in a standard and efficient way, psychologists have developed many scales and different taxonomic structures of personality traits, also in response to specific aims (e.g. to obtain the grade of adaptability to a situation). A currently prominent model of personality traits is the so-called Five Factor Model, according to which the following five domains—identified by the © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 189–194, 2024. https://doi.org/10.1007/978-3-031-49062-0_21
190
F. Bassi et al.
acronym OCEAN—represent the main biologically based and cross-culturally replicable individual differences: Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism. These domains also match the so-called Big Five psycholexical factors, which have been condensed from initial large lists of terms generally used to describe an individual’s stable characteristics and have been replicated in Western as well as non-Western languages [1]. Both the Five Factor Model and the Big Five model describe hierarchical structures of individual differences, with upper- as well as lowerlevel scales and several questionnaires have been developed in accordance with these models. In this paper, we used the IPIP-NEO-120 questionnaire, developed by Johnson in 2014 [2], a measure composed of 120 items (4 for each of six sub-dimension of each of the Big Five factors) that gives a general overview of the personality. On the other hand, handwriting is widely studied both to examine its learning and its alterations due to dysgraphia as well as to tremor or other pathologies [3–5]. The approaches in evaluating handwriting can be divided in either graphical or kinematic features analysis. The first makes use of easily recognizable components as the lowercase letter “t”, lowercase letter “f”, space between the lines, baseline, word slant, connecting words/letters and writing pressure [6]. Thanks to the use of special digital tablets combined with pen capable of writing on ordinary sheets of paper, it is also possible to extract kinematics-related parameters from handwriting during the execution of specific tasks. The variables are calculated on the whole exercise (e.g. total length, duration on sheet or in air, mean velocity or mean pressure, etc.) or as parameters evaluated either on single components (tracts between two successive pen lift) or on single strokes (tracts between two successive minima of the curvilinear velocity) and averaged on the whole test [4, 5, 7]. However, although graphology hypothesizes a link between writing and personality, and there are some studies that analyse it, to our knowledge, only one study investigated connections between personality traits and handwriting and drawing features recorded through a digitizing tablet able to measure speed, pressure applied to the sheet, dimension and inclination of components [6]. The study showed that some personality traits can be revealed by handwriting/drawing features extracted from components. The aim of this paper is to extend these results identifying possible correlations between personality domains of the Big Five model and other kinematic parameters related to motor programming of writing.
2 Materials and Methods 2.1 Subjects and Tasks A total of 61 healthy young subjects were enrolled in this study, 32 females and 29 males, aged between 21 and 37 (mean 24 ± 3.3 years). The handwriting was recorded during three different tasks: cursive copying in a fast (F) or accurate (A) way of the Italian sentence “In pochi giorni il bruco diventò una bellissima farfalla che svolazzava sui prati in cerca di margherite e qualche quadrifoglio” (In a few days the caterpillar became a beautiful butterfly that fluttered over the meadows in search of daisies and the occasional four-leaf clover), and the fast reproduction in cursive lowercase letters of “lelele” (lelele) sequence, without pen lifting for 30 s.
Relationship Between Personality and Kinematic Parameters of Handwriting
191
Furthermore, the subjects were asked to answer the 120 questions of the IPIP-NEO120 test available at https://bigfive-test.com/, which gives a score between 24 and 120 for each of the five factors. 2.2 Data Acquisition and Analysis The handwriting tasks were acquired with a graphics tablet (WACOM Intuos® 3) using an ink pen. A sheet of paper was placed over the tablet to provide visual feedback to the subject during the experiment. The data acquisition system sampled at 200 Hz with a spatial resolution of 0.02 mm the position (x y coordinates), the altitude, the azimuth, and the pressure of the pen. From the acquired data, strokes (segments between two successive minima of the curvilinear velocity) related to motor programming of writing were identified, and the following kinematic parameters were calculated on each stroke and averaged on the whole test: pressure, length, duration, horizontal, vertical, and curvilinear peak and mean velocities. Moreover, on the whole trace were computed the total time spent to complete the exercise, the total length of written space, the mean curvilinear velocity, the total duration of pen lift and the number of strokes. Recordings in which the subjects did not write in cursive were excluded. In order to find possible correlations between each of the 14 handwriting parameters and each of the five personality factors, a linear regression between each pair of variables was computed. Only pairs with a significant linear relation (p < 0.05), were considered. All the data processing was done on MATLAB® .
3 Results Figure 1 shows some examples of the relations between the kinematic variables and the four factors that show a significant linear relation. For Neuroticism, there is no significant relation with any kinematic parameters. For all the relationships, it is possible to observe a pronounced inter-subject variability reflexed in the determination coefficient values, R2 , of the linear regression (Table 1). Table 1 presents all the significant (p < 0.05) linear relations found between the Big Five factors and the kinematic parameters in the different tasks, with the corresponding number of subjects considered. Fourteen relations were found, mainly concerning velocity, pressure, length of written tracts, number of strokes and pen lift duration. In particular, increase of Openness was correlated with decrease of number of strokes and increment of mean curvilinear velocity and peak horizontal velocity of strokes, in the Accurate task. Increase of Agreeableness was associated to decrease of pressure in both Accurate and Fast tasks. Increase of Conscientiousness was linked to decrease in velocities (mean and peak curvilinear, stroke peak horizontal in Fast task and stroke peak horizontal in lelele task) and lengths (stroke mean in Fast and lelele tasks, total written in Fast task). Finally, increase of Extraversion was correlated to increase in number of strokes as well as in decrease of the duration of pen lift, in lelele task.
192
F. Bassi et al.
Fig. 1. Examples of significant linear regression between pairs of personality traits and kinematic parameters in the three writing tests.
4 Discussion The obtained results are not comparable with previous researches [8] because of different experimental protocols and features examined. Given that the comparison is done between objective measurements and self-report scores, an elevated variability is to be expected. Despite this, the study made it possible to highlight the existence of a series of relations existing between some kinematic characteristics, especially linked to motor programming, and almost all personality traits. In particular, considering Openness, all the variables significantly related to were obtained during the Accurate task. Higher value of Openness corresponds to higher writing velocity and lower fragmentation that can be explained remembering that higher Openness generally reflects a higher cultural level and presumably reflect being used to read and write more. Agreeableness presented a significant relation with the mean pressure in both Accurate and Fast exercises, decreasing when pressure increases, thus suggesting that less prosocial, other-oriented and empathic individuals are more incisive in their writing style. About Conscientiousness, several parameters related to velocities and lengths showed a significant inverse relation in Fast and lelele exercises. Greater Conscientiousness corresponds to lower velocity and less space utilized to write the same number of letters during tasks in which high velocity is required, that is, smaller size of the letters. Such a finding is in line with a typical profile of higher conscientiousness individuals which tend to control their cognitive and behavioural impulses, being more self-disciplined, controlled and with planned behaviour.
Relationship Between Personality and Kinematic Parameters of Handwriting
193
Table 1. Significant relations between personality traits and kinematic features. For each relation the corresponding task, R2 , m e q parameters of the linear regression are reported. Personality trait
Task
Feature
R2
Openness
A
Number of strokes
0.08
−1.40
A
Mean curvilinear velocity
0.09
A
Stroke peak horizontal velocity
Agreeableness
A
Conscientiousness
Extraversion
m
q
# of subjects
420
58
0.14
16
58
0.09
0.15
11
58
Mean pressure 0.12
−2.10
380
58
F
Mean Pressure 0.14
−3.00
460
40
F
Peak curvilinear velocity
0.13
−0.32
69
40
F
Stroke mean length
0.14
−0.03
6.5
40
F
Mean curvilinear velocity
0.11
−0.23
54
40
F
Total written length
0.10
−7.00
1600
40
F
Stroke peak horizontal velocity
0.20
−0.31
55
40
lelele
Stroke mean length
0.08
−0.03
8.3
61
lelele
Stroke peak horizontal velocity
0.10
−0.24
50
61
lelele
Number of strokes
0.08
0.35
132
61
lelele
Total duration of pen lift
0.13
5100
61
−46.0
Finally, Extraversion showed significant direct relation with the number of strokes and inverse link with the total duration of pen lift in the lelele task corresponding to a higher fragmentation and a shorter programming time of writing. Such a finding could reflect Extraversion as a higher-order domain including expression of impulses, activity and excitement-seeking.
194
F. Bassi et al.
From a psychological point of view, the Big Five characteristics are not totally independent from each other, so considering a simple regression between a handwriting component and a single personality trait could limit the reliability of results. In any case this preliminary study opens up the possibility of being able to use the kinematic analysis of writing also in the psychological field.
5 Conclusions Digital technologies enable the extraction of kinematic parameters not easily identifiable before. This means that novel correlations between them and other factors like for example the personality tracts can be discovered, increasing the assessment tools available in the psychological field. Aim of the present study was to determine possible correlations between personality traits and kinematic parameters of handwriting, in order to associate objective values to traits in a simple and economical way. The study, conducted on 61 young healthy subjects, outlined some relation between four of the five psychological traits and kinematic parameters of handwriting that mostly deal with velocity, fragmentation, pressure, size of letters and pen lift duration, giving a possible psychological explanation. The results of the study are limited by the great variability present due in part to the small number of subjects who took part in the research. The next step will therefore be to increase the sample size.
References 1. Tupes, E.C., Christal, R.E.: Recurrent personality factors based on trait ratings. J. Pers. 60, 225–251 (1992). https://doi.org/10.1111/j.1467-6494.1992.tb00973.x 2. Johnson, J.A.: Measuring thirty facets of the five factor model with a 120-item public domain inventory: development of the IPIP-NEO-120. J. Res. Pers. 51, 78–89 (2014). https://doi.org/ 10.1016/j.jrp.2014.05.003 3. Gargot, T., et al.: Acquisition of handwriting in children with and without dysgraphia: a computational approach. PLoS One 15, e0237575 (2020). https://doi.org/10.1371/journal.pone.023 7575 4. Smits, E.J., et al.: Standardized handwriting to assess bradykinesia, micrographia and tremor in Parkinson’s disease. PLoS One 9, e97614 (2014). https://doi.org/10.1371/journal.pone.009 7614 5. Alonso-Martinez, C., Faundez-Zanuy, M., Mekyska, J.: A comparative study of in-air trajectories at short and long distances in online handwriting. Cogn. Comput. 9, 712–720 (2017). https://doi.org/10.1007/s12559-017-9501-5 6. Gavrilescu, M., Vizireanu, N.: Predicting the Big Five personality traits from handwriting. J Image Video Proc 2018, 57 (2018). https://doi.org/10.1186/s13640-018-0297-3 7. Mergl, R., et al.: Kinematical analysis of handwriting movements in depressed patients. Acta Psychiatr. Scand. 109, 383–391 (2004). https://doi.org/10.1046/j.1600-0447.2003.00262.x 8. Esposito, A., Amorese, T., Buonanno, M., Esposito, A.M., Faundez-Zanuy, M., LikformanSulem, L., Riviello, M.T., Troncone, A., Cordasco, G.: Handwriting and drawing features for detecting personality traits: an analysis on big five sub-dimensions. Acta Polytechnica Hungarica 19, 65–84 (2022). https://doi.org/10.12700/APH.19.11.2022.11.4
Loneliness and Heart Rate in Older Adults Raquel Cervigón1(B) , Samuel Ruipérez-Campillo2 , José Millet3 , and Francisco Castells3 1 Universidad de Castilla-La Mancha, Cuenca, Spain
[email protected]
2 Department of Medicine, Stanford University, Stanford, USA 3 Universidad Politécnica de Valencia, Valencia, Spain
Abstract. The purpose of the study is to better understand the complex nature of loneliness in older adults and the potential contributing factors that may impact their sense of connection and well-being. The study utilized a mixed-methods approach, combining quantitative measures such as heart rate monitoring with qualitative data collected through interviews and surveys. The findings suggest that loneliness in older adults may be influenced by multiple factors, including their level of education, resilience, and empathy and incidence in spontaneous heart rate variations. Results highlight the importance of empathy in promoting social connectedness and reducing feelings of loneliness in older adults, may have implications for developing targeted interventions aimed at reducing loneliness and improving the well-being of older adults. Keywords: Loneliness · Older people · Heart rate · Psychological factors
1 Introduction The importance of psychosocial factors, has gained significant attention in the study of cardiovascular health. The cardiovascular system is sensitive to changes in emotional states can impact the development of cardiac biomarkers, which are measurable indicators of heart disease risk. Understanding the interplay between psychosocial factors and cardiac signals may provide valuable insights into the development and management of cardiovascular diseases. Loneliness is a subjective feeling of social isolation or lack of companionship. It has been linked to a number of negative health outcomes, including heart disease. One study found that loneliness was associated with higher levels of C-reactive protein (CRP), a biomarker of inflammation that has been linked to heart disease [1, 2]. Other studies have found similar associations between loneliness and other cardiac biomarkers, such as interleukin-6 and fibrinogen [3]. Resilience is the ability to bounce back from adversity or stress. It has been linked to better physical and mental health outcomes. Several studies have investigated the relationship between resilience and cardiac biomarkers. One study found that higher levels of resilience were associated with lower levels of CRP and IL-6 [4] (Steptoe et al., 2008). Another study found that resilience was associated with lower © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 195–203, 2024. https://doi.org/10.1007/978-3-031-49062-0_22
196
R. Cervigón et al.
levels of cortisol, a hormone that has been linked to heart disease [5]. Resilience, on the other hand, has been linked to decreased sympathetic nervous system activation and increased parasympathetic nervous system activity, which may have a protective effect on the cardiovascular system. Studies have found that resilient individuals have lower levels of pro-inflammatory cytokines and cortisol, and higher levels of anti-inflammatory cytokines, which may help protect against cardiovascular disease [6, 7]. Purpose in life has been associated with better cardiovascular health outcomes, including lower incidence of cardiovascular disease, lower risk of stroke, and lower mortality rates [8, 9]. It has also been linked to improved self-reported health, better sleep quality, and lower levels of inflammation [10, 12]. Empathy, on the other hand, has been linked to increased heart rate variability and cardiac coherence. Cardiac coherence is a measure of the synchronization between heart rate variability and breathing rate, and it is associated with improved cardiovascular health [13]. Studies have shown that empathy training can increase heart rate variability and cardiac coherence, suggesting that empathy may have protective effects on cardiovascular health [14, 15]. Cardiac signals, such as heart rate variability (HRV) and heart rate (HR), are important markers of cardiovascular health and have been linked to psychosocial factors. HRV reflects the balance between sympathetic and parasympathetic nervous system activity and is a marker of cardiac autonomic regulation. HR is a simple and easily obtainable measure of cardiac function that has been linked to various psychosocial factors. Understanding the link between psychosocial factors and cardiac signals may provide valuable insights into the pathogenesis of CVD and the development of novel interventions for its prevention and management. This paper aims to analysis the relationship between loneliness, resilience, empathy, purpose in life, and cardiac signals, including HRV and HR, through a study using wearables for recording physiological measures and administering psychological questionnaires in older adults.
2 Materials and Methods This is a cross-sectional study that analyses a significant sample of participants from 55 to 94 years old living in the city of Cuenca. The data were collected at doctor’s appointments or during their visit to social services. Consecutive sampling was carried out and justifies a sample size of 210 people in order to make an error of 5%, with a confidence level of 95%. From this number, those with cognitive dysfunction should be excluded. The questionnaires were handed out to eligible patients who agreed to participate in the study by signing the informed consent form. During the questionnaire they were fitted with the Empatica E4 wristband which records the instantaneous pulse signal, has an accelerometer, a temperature sensor and records the electrical activity of the skin. Inclusion criteria were age less than 55 years and being a patient of the cardiology service and living in Cuenca and exclusion criteria were cognitive impairment and/or cardiac arrhythmia (Table 1).
Loneliness and Heart Rate in Older Adults
197
Table 1. Characteristics of participants. Parameters
Gender Women
Age
Civil state
Education
Total Men
55–64
7
36
43 (20%)
65–74
20
59
79 (38%)
75–84
18
49
67 (32%)
85–94
6
15
21 (10%)
Single
6
23
29 (10%)
Married
39
69
108 (10%)
Widower
4
50
54 (10%)
Divorced
2
17
19 (10%)
Primary studies
99
24
123 (58%)
Secondary studies
26
9
35 (17%)
University studies
34
18
52 (25%)
The Meaning of Life Questionnaire (MLQ) validated by Michael F. Steger was administered [16]. The validated Spanish version assessed two dimensions of meaning in life by means of 10 items rated on a seven-point scale. The MLQ was rated on a seven-point scale from “Absolutely true” to “Absolutely false”. The Presence of Meaning subscale measures the degree of fullness of meaning that respondents feel in their lives. In addition, the Search for Meaning subscale measures the degree to which respondents are engaged and motivated in their efforts to find meaning or deepen their understanding of meaning in their lives. Loneliness was also assessed using the De Jong Gierveld (DJG) loneliness scale. The 11-item DJG scale, validated in Spanish, includes a 6-item emotional subscale (negatively worded) and a 5-item social subscale (positively worded). Recent studies suggest that the DJG scale may be a better choice in research specifically involving middle-aged and older adults [17]. Conceptually, isolation can be thought of as an objective lack of social contact or interaction, and loneliness as the subjective experience of lack of companionship, social interaction and bonding - both social and emotional - while isolation is an “objective” lack of bonding due to isolation at home, with little or no social interaction with others or access to outside activities. One of the most useful and internationally used measures is the Empathy Quotient (EQ) [18]. The EQ is a short, accessible questionnaire that is quick and easy to correct. It was specifically designed to be applied in clinical settings to assess lack of empathy as a psychopathological trait, as well as to detect subtle differences in the level of empathy in the general population and to capture gender differences in this construct. Nevertheless, the main objective is the analysis of biomarkers to assess the relationship between loneliness and cardiac response. The biomarkers of the patients who will be part of the study contain skin conductance, blood volume pulse, accelerometers = physical activity, HR and skin temperature. Empatica E4 Wristband may be specially intended for researchers who want to understand human emotions. This device is equipped with:
198
– – – –
R. Cervigón et al.
PPG sensor: measures heart rate variability through blood volume pulse EDA sensor: measures electrodermal activity. 3-axis accelerometer: measures acceleration and movement Infrared thermopile: measures skin temperature
2.1 Signal Processing Analysis The E4 wristband is designed to measure various physiological signals such as the Inter-Beat Interval (IBI) and Blood Volume Pulse (BVP). The IBI refers to the time interval between successive heartbeats and can be used to calculate heart rate variability (HRV), which is a measure of the variation in time between heartbeats. HRV is useful for assessing the health of the autonomic nervous system and has been linked to various health outcomes. BVP refers to the changes in blood volume in the peripheral blood vessels with each heartbeat. This signal can be used to measure changes in blood flow, which is useful for assessing changes in skin blood perfusion and sweat gland activity. Signal processing is a crucial step in analyzing the data collected from the E4 wristband. The recorded physiological signals need to be filtered and processed to remove noise and artifacts before any meaningful analysis can be performed. Various signal processing techniques, such as filtering, detrending, and artifact removal, can be applied to the raw data to extract relevant information. Empatica recordings typically include two types of signals: Inter-Beat Interval (IBI) and Blood Volume Pulse (BVP). IBI is the time interval between two consecutive heartbeats, while BVP is a measure of the blood volume changes in the fingertip, reflecting changes in blood flow with each heartbeat. The following are the steps to process the recordings: 1. Pre-processing: This involves removing any artifacts and noise from the raw signals, such as motion artifacts, baseline drift, and electrode movement. This can be done using various signal processing techniques, such as filtering, detrending, and artifact removal. 2. Peak detection: The IBI signal is obtained by detecting the time between consecutive heartbeats. The first step in obtaining IBI is to identify the peak of the BVP signal, which corresponds to the time when the blood volume in the blood vessels is at its maximum. This peak is detected using a peak detection algorithm. 3. Inter-beat Interval Calculation: Once the peaks of the BVP signal are detected, the IBI can be calculated as the time interval between two consecutive peaks. This interval is measured in milliseconds. 4. Feature extraction: Various features can be extracted from the IBI signal, such as the mean IBI, standard deviation of IBI, frequency-domain features, and time-domain features. This involves computing various statistical parameters from the IBI and BVP signals, such as: IBImax, IBImmedian, SDNN (which reflects variability in heart rate), SDANN (standard deviation of the average IBI over 5-min intervals, which reflects the longer-term variability in thein, IBImean, IBI heart rate), NNx (number of pairs of adjacent IBIs that differ by more than x milliseconds, where x is a predefined threshold), RMSSD (which reflects the short-term variability in the
Loneliness and Heart Rate in Older Adults
199
heart rate and SDNNi (standard deviation of the IBI values after removing artifacts and outliers). 5. Analysis and interpretation: The extracted features can be analyzed and interpreted to understand various aspects of the cardiovascular system, such as heart rate variability, autonomic nervous system activity, and stress levels. Canonical Correlation Analysis (CCA) was applied to investigates the relationships between two sets of variables, emotional and social loneliness with empathy, resilience, meaning in life and heart rate measurements. The mathematical definition of CCA is as follows: We have two sets of variables X and Y, each containing p and q variables, respectively. Let X be an n×p matrix, where each row represents an observation of the p variables, and let Y be an n×q matrix, where each row represents an observation of the q variables. The goal of CCA is to find linear combinations of X and Y that are maximally correlated with each other. These linear combinations are called canonical variables. The canonical variables are computed as follows: Let a and b be column vectors of coefficients such that a = b = 1. The first canonical variable pair (u1, v1) is defined as the pair of linear combinations of X and Y that maximizes the correlation between u1 = X a and v1 = Y·b, subject to the constraint that a ·X ·X·a = b ·Y ·Y·b = 1. The second canonical variable pair (u2, v2) is defined as the pair of linear combinations of X and Y that maximizes the correlation between u2 = X·c and v2 = Y·d, subject to the constraint that c ·X ·X·c = d ·Y ·Y·d = 1, and c ·X ·Y·u1 = d · Y ·X·u1 = 0. The process is repeated to obtain k pairs of canonical variables. The kth canonical variable pair (uk, vk) is defined as the pair of linear combinations of X and Y that maximizes the correlation between uk = X·ak and vk = Y·bk, subject to the constraint that ak’·X ·X·ak = bk’·Y ·Y·bk = 1, and ak’·X ·Y·u1 = bk’·Y ·X·u1 = · · · = ak’·X ·Y·u(k-1) = bk’·Y ·X·u(k-1) = 0. The canonical correlations are the square roots of the canonical variable correlation coefficients, and they represent the strength of the linear relationship between the canonical variables. The canonical correlations range from 0 to 1, with larger values indicating stronger relationships. One-way ANOVA (Analysis of Variance) was also tested when the level of loneliness was categorized into three stages (not lonely, moderate lonely and severe lonely). It is a statistical technique used to compare the means of two or more groups, in this case three groups. It tests the null hypothesis that there is no significant difference between the means of the groups against the alternative hypothesis that at least one group mean is different from the others.
3 Results CCA conducted to examine the relationship between two sets of variables (Table 2), set 1: social loneliness and emotional loneliness and set 2: NNx (high variations of inter-beat intervals) empathy and resilience. The analysis aimed to determine the extent to which
200
R. Cervigón et al.
two sets of variables were related to each other, as well as the specific variables that were most strongly correlated. Table 2. Canonical correlations Correlation
Canonical correlation
Eigenvalue
Percent of variance
CanCorr1
0.50
0.31
79.10%
CanCorr2
0.28
0.09
20.89%
Non statistical significant differences were found in age, gender or other demographic variables. The first canonical variate explains 79.10% of the total variance, while the second canonical variate explains an additional 20.89% of the total variance. The second canonical correlation had the lowest eigenvalue and the lowest correlation coefficients. Together, these three canonical correlations explain 100% of the variance in the relationship between the two sets of variables. Results show a significant difference on the set of dependent variables representing emotional loneliness (Wilks Lambda = 0.77; F(6, 410) = 9.43, p < 0.001, with resilience, presence of meaning in life and short time variability of heart rate (NNx). Nevertheless, when we analyzed as independent variable, social loneliness, NNx was not statistical significant. Furthermore, when the loneliness variable was categorized into three stages, an analysis of variance (ANOVA) indicated that the education variable was significant. This suggests that education level may have an impact on loneliness in the population studied as well as the variable resilience. A one-way ANOVA was conducted to determine if there were any significant differences in the mean scores on resilience, level of studies, civil state, empathy, meaning in life and heart rate measurements between 3 leves of loneliness: not lonely, moderate lonely and severe lonely. The variables where the difference was higher are shown in the Figs. 1 and 2, nevertheless level of studies, civil state, empathy, meaning in life and resilience showed statistical differences across different groups of loneliness. The assumption of homogeneity of variances was met (In empathy and resilience variables, with Levene’s test F[2, 207] = 0.951, p = 0.39 and F[2, 207] = 0.41, p = 0.66, respectively), indicating that the groups had equal variances. The results of the ANOVA revealed a significant main effect of group on resilience and empathy scores (F[2, 207] = 9.03, p < 0.001 and F[2, 207] = 1.91, p = 0 .009, respectively).
4 Discussion Loneliness is a significant and growing concern among older adults, and research has demonstrated that it can have negative effects on a range of health outcomes, including CVD [19, 20]. While previous studies have found that loneliness is associated with increased resting heart rate, recent evidence suggests that it may also be negatively correlated with heart rate variability (HRV), a measure of the beat-to-beat changes in heart rate that reflects the balance between sympathetic and parasympathetic nervous
Loneliness and Heart Rate in Older Adults
201
Fig. 1. Resilience in relation to loneliness
Fig. 2. Presence of meaning in life in relation to loneliness
system activity [21, 22]. HRV refers to the variation in time between heartbeats and is considered an indicator of cardiovascular health. Lower HRV has been linked to increased risk of cardiovascular disease and mortality. This relationship may be due to the effects of loneliness on the autonomic nervous system, which can lead to dysregulation of heart rate and increased risk of cardiovascular disease [3]. Empathy refers to the ability to understand and share the feelings of others. Research has found that loneliness in older adults is negatively correlated with empathy, suggesting that loneliness may impair social cognition and emotional processing. This relationship
202
R. Cervigón et al.
may be due to the fact that loneliness can lead to social withdrawal and reduced opportunities for social interaction, which can in turn impair empathic abilities [23]. Furthermore, loneliness has also been negatively correlated with empathy in older adults. Meaning in life refers to a sense of purpose and direction, and has been found to be an important predictor of mental health and well-being. Research has found that loneliness in older adults is negatively correlated with meaning in life, suggesting that loneliness may lead to a sense of existential emptiness and disconnection [24]. Resilience refers to the ability to adapt and cope with adversity, and has been found to be an important factor in maintaining mental and physical health. Research has found that loneliness in older adults is negatively correlated with resilience, suggesting that loneliness may impair coping abilities and increase vulnerability to stress [25]. Finally, loneliness has been negatively correlated with level of education in older adults. Education is an important predictor of health and well-being in later life, and lower levels of education are associated with a range of negative outcomes [26]. Present results suggest that loneliness may be a risk factor for lower educational attainment in older adults. In conclusion, the evidence suggests that loneliness in older adults is negatively correlated with a range of psychological and physiological outcomes, including heart rate changes, empathy, meaning in life, resilience, and level of education. These findings highlight the importance of addressing loneliness in older adults as a means of promoting health and well-being in later life. Interventions aimed at reducing loneliness in older adults may have important benefits for both physical and psychological health. In summary, our findings suggest that high spontaneous variations in heart rate, empathy and resilience are correlated with both loneliness, social and emotional. Moreover, education and civil state can affect in a significant factor in predicting levels of loneliness when classified into three stages. These results contribute to our understanding of the complex factors that influence loneliness and highlight the importance of considering both physiological and socio-demographic variables when studying this phenomenon. Acknowledgements. This work was supported in part by grants from the Spain’s Foundation of Superior Council of Scientific Investigations (POCTEP 0551-PSL-6-E).
References 1. Hawkley, L.C., Burleson, M.H., Berntson, G.G., Cacioppo, J.T.: Loneliness in everyday life: cardiovascular activity, psychosocial context, and health behaviors. J. Pers. Soc. Psychol. 85(1), 105–120 (2003) 2. Cacioppo, J.T., Hawkley, L.C.: Social isolation and health, with an emphasis on underlying mechanisms. Perspect. Biol. Med. 46(3Suppl), S39-52 (2003) 3. Hawkley, L.C., Thisted, R.A., Masi, C.M., Cacioppo, J.T.: Loneliness predicts increased blood pressure: 5-year cross-lagged analyses in middle-aged and older adults. Psychol. Aging 25(1), 132–141 (2010) 4. Steptoe, A., Diez Roux, A.V.: Happiness, social networks, and health. BMJ 337, a2781 (2008) 5. Goldman-Mellor, S., Hamer, M., Steptoe, A.: Early-life stress and recurrent psychological distress over the lifecourse predict divergent cortisol reactivity patterns in adulthood. Psychoneuroendocrinology 37(11), 1755–1768 (2012)
Loneliness and Heart Rate in Older Adults
203
6. Teche, S.P., et al.: Resilience to traumatic events related to urban violence and increased il10 serum levels. Psychiatry Res. 250, 136–140 (2017) 7. Seiler, C., Holmes, S.: Multivariate heteroscedasticity models for functional brain connectivity. Front. Neurosci. 11, 696 (2017) 8. Hill, P.L., Turiano, N.A.: Purpose in life as a predictor of mortality across adulthood. Psychol. Sci. 25(7), 1482–1486 (2014) 9. Lau, S.C.L., Tabor Connor, L., Baum, C.M.: Motivation, physical activity, and affect in community-dwelling stroke survivors: an ambulatory assessment approach. Ann Behav Med (2023) 10. Kim, H.-G., Cheon, E.-J., Bai, D.-S., Lee, Y.H., Koo, B.-H.: Stress and heart rate variability: a meta-analysis and review of the literature. Psychiatry Investig. 15(3), 235–245 (2018) 11. Baylis, D., Bartlett, D.B., Patel, H.P., Roberts, H.C.: Understanding how we age: insights into inflammaging. Longev Healthspan 2(1), 8 (2013) 12. Shaffer, F., McCraty, R., Zerr, C.L.: A healthy heart is not a metronome: an integrative review of the heart’s anatomy and heart rate variability. Front. Psychol. 5, 1040 (2014) 13. Lehrer, P., et al.: Negative mood and alcohol problems are related to respiratory dynamics in young adults. Appl. Psychophysiol. Biofeedback 38(4), 273–283 (2013) 14. Fredrickson, B.L., et al.: Positive emotion correlates of meditation practice: a comparison of mindfulness meditation and loving-kindness meditation. Mindfulness (N Y) 8(6), 1623–1633 (2017) 15. Steger, M.F., Frazier, P., Oishi, S., Kaler, M.: The meaning in life questionnaire: assessing the presence of and search for meaning in life. J. Couns. Psychol. 53(1), 80 (2006) 16. Tomas, J.M., Pinazo-Hernandis, S., Donio-Bellegarde, M., Hontangas, P.M.: Validity of the De Jong Gierveld loneliness scale in Spanish older population: competitive structural models and item response theory. Eur. J. Ageing 14(4), 429–437 (2017) 17. Baron-Cohen, S., Wheelwright, S.: The empathy quotient: an investigation of adults with Asperger syndrome or high functioning autism, and normal sex differences. J. Autism Dev. Disord. 34(2), 163–175 (2004) 18. Hawkley, L.C., Cacioppo, J.T.: Loneliness matters: a theoretical and empirical review of consequences and mechanisms. Ann. Behav. Med. 40(2), 218–227 (2010) 19. Victor, C.R., et al.: The prevalence and predictors of loneliness in caregivers of people with dementia: findings from the ideal programme. Aging Ment. Health 25(7), 1232–1238 (2021) 20. Conde-Sala, J.L., Garre-Olmo, J. Calvó-Perxas, L. Turró-Garriga, O., Vilalta-Franch, J., Lopez-Pousa, S.: Causes, mortality rates and risk factors of death in community-dwelling Europeans aged 50 years and over: results from the survey of health, ageing and retirement in Europe 2013–2015. Arch. Gerontol. Geriatr. 89, 104035 (2020) 21. Fuller-Rowell, T.E., Williams, D.R., Love, G.D., McKinley, P.S., Sloan, R.P., Ryff, C.D.: Race differences in age-trends of autonomic nervous system functioning. J. Aging Health 25(5), 839–862 (2013) 22. Cacioppo, J.T., Hawkley, L.C., Thisted, R.A.: Perceived social isolation makes me sad: 5-year cross-lagged analyses of loneliness and depressive symptomatology in the Chicago health, aging, and social relations study. Psychol. Aging 25(2), 453–463 (2010) 23. Lafrenière, M.-A.K., Sedikides, C., Van Tongeren, D.R., Davis, J.: On the perceived intentionality of self-enhancement. J. Soc. Psychol. 156(1), 28–42 (2016) 24. Hawkley, L.C., Preacher, K.J., Cacioppo, J.T.: Loneliness impairs daytime functioning but not sleep duration. Health Psychol. 29(2), 124–129 (2010) 25. Perissinotto, C.M., Stijacic Cenzer, I., Covinsky, K.E.: Loneliness in older persons: a predictor of functional decline and death. Arch Intern Med 172(14), 1078–1083 (2012)
Filters for Electrocardiogram Signal Processing: A Review Elma Kandi´c1
and Lejla Gurbeta Pokvi´c2(B)
1 Faculty of Electrical Engineering, University of Sarajevo, Sarajevo, Bosnia and Herzegovina 2 Medical Device Inspection Laboratory VERLAB, Sarajevo, Bosnia and Herzegovina
[email protected]
Abstract. Recording an Electrocardiogram (ECG) signal is a difficult task in the field of biomedical engineering. The ECG signal reflects the electrical activity of the heart muscle and is important in diagnosing heart conditions. However, the signal is often contaminated with various types of noise during processing, such as muscle noise, power line interference, baseline wandering, and motion artifacts. It is crucial to separate the desired signal from these noise sources to ensure accurate diagnosis. This article examines the challenges associated with ECG preprocessing filters in the last five years. Keywords: ECG · Filter · Signal processing
1 Introduction The outputs from biosensors are analog signals, which are sent to the analog processing and digital conversion block. There, the signals are amplified, filtered, conditioned, and converted to digital form. The signal that is often used in these modifications is the electrocardiogram (ECG). In the simultaneous contraction of both ventricles blood is forced from the heart into the pulmonary artery from the right ventricle and into the aorta from the left ventricle. The electrocardiogram is an electrical measure of the sum of these ionic changes within the heart. Throughout the data acquisition procedure, it is critical that the information and structure of the original biological signal of interest as faithfully preserved. ECG data is later processed to determine arrhythmic activity and other important diagnostic characteristics. ECG signals are essential to diagnose and analyze cardiac disease, because ECG signals record the cardiac electrical activity, which conveys important pathological information about the human heart’s condition. By analyzing the characteristics of ECG, doctors are able to judge whether the heart situation is normal or not, and know what troubles the heart confronts with. Since these signals are often used to aid the diagnosis of pathological disorders, the procedures of amplification, analog filtering or A/D conversion should not generate misleading or untraceable distortions. Distortions in a signal measurement could lead to a delay in the initiation of appropriate medical treatment or to an improper diagnosis. An ECG typically contains unwanted interference or noise. Such interference has the detrimental effect of obscuring relevant information that may © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 204–217, 2024. https://doi.org/10.1007/978-3-031-49062-0_23
Filters for Electrocardiogram Signal Processing: A Review
205
be available in the measured signal. Interference noise occurs when unwanted signals are introduced into the system by outside sources. It is introduced by power lines (50 or 60 Hz), fluorescent lights, AM/FM radio broadcasts, computer clock oscillators, laboratory equipment, cellular phones, and so forth. Even the action potentials from nerve conduction in the patient generate noise at the sensor/amplifier interface. Also, ECG measurements from the heart can be affected by bioelectric activity from adjacent muscles. A measurement ECG electrode can pick up extraneous signals from the muscles, lungs, and even from the internal electronics of the recording devices. An ECG has very small magnitudes, approximately in the millivolts. Filters are often used to remove noise from a signal, typically through the use of frequency-domain analysis to design the filter. Appropriate filtering allows one to clean up the signal, thus improving its quality of signal and the diagnostic reliability in clinical settings. Noise filtering is the fundamental step in the processing of the ECG signal. Alternating current (AC) source from a power supply introduces the PLI noise, which is a major noise to be removed at the initial stage of processing steps. Based upon the country region, the signal has a frequency of about 50/60 Hz. The main reasons behind such type of noise are the stray effect of alternating current field because of loops in the electricity wires, disengaged electrodes, electromagnetic interference due to power supply, improper grounding of ECG equipment, or heavy current load due to other equipment in the room. A low-frequency noise called baseline wander noise also occurs during ECG recording. It has the range of 0.15 to 0.3 Hz. This noise occurs due to the breathing process of that person and forces the ECG signals to shift in the baseline. The other probable causes may be due to the movement of cables during the recording of the ECG signal or due to unclean lead electrodes/wires, or due to loosen electrode connection. In addition to the heart, muscle contraction contributes to the electromyogram (EMG) noise due to depolarization and repolarization waves generated from muscle contraction near the electrodes. Another type of noise is contact noise. It is caused by the heart’s position in relation to the electrode’s variance. Electrode–skin impedance variation is the mechanism responsible for baseline disturbances. An artifact called electrode motion artifact occurs due to the movement of electrodes. The subject’s vibrations, movement, or breathing usually contribute to motion artifacts. Due to very slow fluctuations in the impedance of the skin electrode, a baseline drift arises at a very low frequency in the ECG signal. This noise cannot be disposed of, but high-quality hardware and a cautious circuit plan can very well reduce it. The main reasons are the connection of electrodes, wires, signal processor/amplifier, and ADC. At the hospitals, nurses and doctors do not pay attention to electrode placement. It results in common mode noise, and therefore 50 Hz filtering must be used.
206
E. Kandi´c and L. G. Pokvi´c
This work attempts to summarize filtering methods and approaches into a complete overview and categorize them into a systemic taxonomy. Therefore, the purpose of this paper is to review the latest achievements in this field in the last 5 years.
2 Methods The aim of this review paper was to provide an analysis of filters used for electrocardiogram (ECG) signal processing. A literature search was conducted in multiple databases, including PubMed, Scopus, IEEE Xplore, and Web of Science, using relevant search terms and keywords such as ECG, signal processing and filter. The search was limited to studies published in English and the search was conducted within the last 5 years. The studies identified in the search were screened based on predefined inclusion and exclusion criteria. Studies were included if they described the use of filters in ECG signal processing, and were published in peer-reviewed journals, conference proceedings, or books. Studies were excluded if they did not meet the inclusion criteria or were published in a language other than English. Data was extracted from each study, including the type of filter used, the characteristics of the ECG signals processed, the performance metrics used to evaluate the filter, and the main findings of the study. The extracted data was organized by themes, The themes included the different types of filters used in ECG signal processing. The extracted data was critically analyzed to identify patterns and trends in the literature, and to draw conclusions about the effectiveness of different types of filters for ECG signal processing. Overall, the methods (Table 1) used in this review paper were designed to provide a rigorous and systematic approach to the literature review process, and to ensure that the analysis was comprehensive, accurate, and unbiased.
3 Results In the study conducted by Venkatesan et al. (2018) [8], author utilized a standardized Least Mean Squares (LMS) adaptive filter with a delayed error in the preprocessing stage to achieve higher speed and a low-latency design with fewer elements, mainly to remove white noise. The results were compared with second-order IIR notch filter (Tables 2 and 3). Huang, Hui, Shiyan Hu, and Ye Sun (2019) [4] considered a new low-distortion adaptive Savitzky-Golay (LDASG) filtering method for ECG denoising based on discrete curvature estimation (see Fig. 1). With comparable noise elimination performance, the standard SG filter has greater distortions at high variation parts, especially at R peaks, than the proposed method. The proposed method is compared with the EMD-wavelet-based method and the non-local means (NLM) denoising method in terms of both noise elimination and signal distortion reduction. For signal distortion reduction, their method outperforms the EMD-wavelet method by reducing MSE by 33.33% and PRD by 18.25%, and outperforms the NLM method by reducing MSE by 50% and PRD by 25.24% (Tables 4 and 5).
Filters for Electrocardiogram Signal Processing: A Review
207
Table 1. A brief overview of the considered filters. Filter/method
Purpose
Comment
The least mean square (LMS) adaptive filter
Removes white noise
The higher speed and a low-latency design
Low-distortion adaptive Savitzky-Golay (LDASG)
Denoises high signal variations with low signal distortion
The remarkable performance of data smoothing
Wavelet without applying thresholding
Identifies the ECG and noise More accurate algorithm with frequency range for zeroing a more efficient estimation of wavelet detail coefficients in the baseline subbands with no ECG coefficients in the frequency content
Combined EMD method
Efficiently removes a pure 50 Hz The better improvement in sinusoidal noise termed as SNR values of the signal power line interference
Adaptive wavelet A new shrinkage function thresholding method (AWT) depending on the identical correlation is utilized to realize smooth transition from the thresholding cutoff to a real wavelet coefficient
The distinctadaptability in the selections of base wavelet, threshold and shrinkage function
DWT with the β-hill climbing
Finds the optimal wavelet parameters that will minimize the mean square error (MSE) between the original and denoised signals
The mean square error is minimised
Bayesian denoising framework
Credited to ECG denoising, segmentation and arrhythmia detection
An adaptive particle weighting strategy. Helpful algorithm, especially in removing baseline drifts caused by MA noises
Synchro-squeezed wavelet transform (SSWT)
Uses for baseline wander correction and powerline interference reduction in electrocardiogram (ECG) signals using empirical wavelet transform (EWT)
A significant improvement in output signal-to-noise ratio
(continued)
The author Jain et al. (2018) [2] designed a robust system for ECG denoising, incorporating EMD algorithm with fractional integral filtering by “Riegmann Liouvelle (RL) and Savitzky–Golay (SG).” It is proved from the results that the EEMD-PSO and EEMD-CS methods give the best performance for denoising attaining maximum SNR and minimum MSE for all types of noises (Table 6).
208
E. Kandi´c and L. G. Pokvi´c Table 1. (continued)
Filter/method
Purpose
Comment
EMD with SWT and a mean Uses the features of EMD as Noticeably enhanced the based filter well as that of the SWT and SNR value of the ECG NLM filter to efficiently remove signals a pure 50 Hz sinusoidal noise termed as power line interference Switching Kalman filter
Extracts fiducial points of ECG signals
The better performance in noisy environments
Kalman filter
Estimates the underlying signal and removes noise and artifacts from the recorded ECG signal
Applicable only for linear systems
Wiener filter
Estimates the optimal filter coefficients for each sample of the signal
The mean square error is minimised
Adaptive notch filters with sharp resolution
Eliminates signals from the stop The specific freq. range and band. Eliminates PLI only one eliminated noise at a time
Table 2. Performance analysis of ECG records using second-order IIR notch filter. Record
SNRinput
SNRoutput
SNRimproved
MSE
RMSE
NMSE
PRD
100
23.35
58.89
33.53
1.0211
1.0105
2.10 × 10−7
4.58 × 10−6
102
19.98
51.41
31.46
1.2233
1.1060
5.09 × 10−7
7.13 × 10−6
105
23.41
58.26
34.84
2.0429
1.4293
3.81 × 10−7
6.17 × 10−6
111
20.25
57.66
37.41
0.6868
0.8251
4.54 × 10−7
6.74 × 10−6
115
23.08
58.01
34.93
3.8366
1.9587
3.16 × 10−7
5.62 × 10−6
To eliminate the white Gaussian noise in the ECG signals. Alyasseri et al. (2017) [28] have suggested combining the DWT with the β-hill climbing technique for suppressing the white Gaussian noise in the ECG signals. Hesar and Mohebbi (2017) [29] have proposed the model based Bayesian denoising framework, which utilizes the DWT based thresholding with the Variational Mode Decomposition (VMD) to lower the noise impact on the ECG signals and then adopts the Marginalized Particle-Extended Kalman Filter (MP-EKF) with the Fuzzy Based
Filters for Electrocardiogram Signal Processing: A Review
209
Table 3. Performance analysis of different ECG records using adaptive LMS filter. Record
SNRinput
SNRoutput
SNRimproved
MSE
RMSE
NMSE
PRD
100
23.35
55.34
29.98
0.0226
0.1503
1.83 × 10−4
1.35 × 10−4
102
19.98
49.08
29.09
0.0063
0.0797
1.04 × 10−4
1.02 × 10−4
105
23.41
56.72
33.31
0.0379
0.1948
2.76 × 10−4
1.66 × 10−4
111
20.25
48.89
28.64
0.0015
0.0385
3.91 × 10−5
6.62 × 10−5
115
23.08
52.29
29.21
0.6189
0.7867
1.9 × 10−3
4.41 × 10−4
Fig. 1. Diagram of the proposed LDASG filter for ECG signal denoising.
Adaptive Particle Weighting (FBAPW) technique to further tackle the noises in the signals. Singh and Sunkaria (2017) [30] have made use of the EWT with the technique of mode subtraction for dealing with different kinds of noises in the ECG signals. SynchroSqueezed Wavelet Transform (SSWT) can also realize the adaptive time-frequency decomposition, which is the goal of EMD. Oliveira et al. (2018) [5] found approach to be superior to normal threshold and notch filtering techniques in removing power-line interference. A new phenomenon called adaptive wavelet thresholding method (AWT) was designed in the paper by He and Tan (2018) [27] for the ECG signal enhancement. By means of cross-relation coefficient and entropy energy relation, the best base wavelet was generated for ECG signal filtering automatically. S. A. Malik, S. A. Parah and G. M. Bhat (2021) [31] concluded that combined denoising capabilities of classical EMD method provided a better improvement in SNR
210
E. Kandi´c and L. G. Pokvi´c Table 4. Results of ECG denoising performance with the SNR level of 0 dB.
ECG records
SNR improvement
MSE
PRD (%)
EMD-w NLM LDASG EMD-w NLM LDASG EMD-w NLM LDASG
#101
9.5
9.05 10.47
0.015
0.017 0.012
33.48
35.29 29.96
#103
7.16
7.83 10.35
0.029
0.026 0.014
43.87
40.61 30.37
#104
8.85
7.79 10.39
0.016
0.021 0.011
36.09
40.88 30.21
#105
9.71
8.22 10.78
0.015
0.022 0.011
32.71
38.83 28.91
#106
6.79
6.16
9.45
0.041
0.048 0.022
45.73
49.23 33.69
7.29
#115
8.06
9.17
0.015
0.061 0.039
39.55
43.16 34.78
#117
13.85
11.24 14.91
0.031
0.056 0.024
20.28
27.41 17.98
9.13
8.23 10.79
0.03
0.04
35.96
39.34 29.41
Average
0.02
Table 5. Comparison of the computation time of different methods (seconds). ECG records
EMD-Wave1et
NLM
LDASG
#101
0.792
0.490
0.550
#103
0.730
0.490
0.531
#104
0.767
0509
0.583
#105
0.815
0.469
0.582
#106
0.754
0.507
0.459
#115
0.789
0.507
0.569
#117
0.779
0.474
0.622
Average
0.775
0.492
0.556
Table 6. Comparison of EMD with fractional integral filtering with other related methods. Method
Random noise
Gaussian noise
PLI noise
SNR (dB)
SNR (dB)
MSE
SNR (d)
MSE
27.72
2.37
1.55 × 10 −4
0.0125
1.34 × 10 −3
8.7892
0.002
6.8665
0.0032
6.355
0.0035
EEMD-GA
8.7892
0.002
5.8147
0.004
9.4462
0.0017
EEMD-CS
10.0809
0.0015
7.97
0.0024
9.4673
0.0017
EEMD-PSO
10.0809
0.0015
7.9776
0.0024
9.4673
0.0017
EMD EMD-norm-2
23.35
MSE
values of the signal in comparison to the method involving only EMD or wavelet based method. The clinical features are preserved and ECG was not compromised (Table 7).
Filters for Electrocardiogram Signal Processing: A Review
211
Table 7. Considered different wavelet transforms. Researches
Filter/method
Noise
A. Alyasseri, A. Khader, A. Al-Betar, and L. M. Abualigah
β-Hill climbing WGN algorithm and wavelet transform
SNRin = 0 dB SNRout = 8.9649 dB
H. D. Hesar and M. Mohebbi
Soft thresholding WGN based on VDM + EKF
SNRin = −1 dB SNRout = 1.395 ± 0.539 dB
O. Singh and R. K. Sunkaria
EWT with mode subtraction
SNRin = 0 dB SNRout = 0.9943 dB
PLI BW
Results (signal #100)
SNRin = 0 dB SNRout = 16.4016 dB Oliveira, Bruno Rodrigues de, et al.
A wavelet-based method without applying thresholding techniques
PLI
SNRimp (avg) = 40.7086 dB
He H and Tan Y
Adaptive wavelet thresholding
PLI, BW and muscle artifact
The AWT has obtained the lowest re and modest values of SNR and RMSE
S. A. Malik, S. A. Parah and G. M Bhat
EMD with SWT and a PLI mean based filter
SNRin = 0 dB SNRIMP = 10.74 dB
Akhbari, Mahsa, et al. (2018) [13] presents a new approach for extracting fiducial points (FPs) of ECG signals by using a switching Kalman filter (SKF). The proposed method is compared with methods based on wavelet transform. For the proposed method, the mean error and the root mean square error across all FPs are 2 ms (i.e. less than one sample) and 14 ms, respectively. The standard deviations are around four to five samples for the onset and offset of waves and around one sample for the peak of waves. The errors and the standard deviation and RMSE values for the SKF are significantly smaller than those obtained using other methods (Table 8). Authors Manju, B. R., and M. R. Sneha (2020) [10] concluded that the Wiener filter is a method of denoising a signal that involves using the spectral properties of the signal and the noise (Fig. 2). The results indicate that the Wiener filter (Fig. 3) produces a higher SNR value, low MSE, and low PRD compared to the Kalman filter (Fig. 2) for all types of noise. The simulation results have shown that Wiener filter is a better filtering technique than Kalman filter in terms of SNR, PSD, MSE, PRD. The inefficient performance of the Kalman filter is due to its restricted application to non-linear systems (Table 9). Chen, Binqiang, et al. (2019) [11] focuses on eliminating PLI from ECG and proposes an Adaptive Notch Filter with Sharp Resolution (ANFwSR) that do not require any
Pon
−0.1 ± 1.5 1.5
0.7 ± 25 25
Pon
23.4 ± 15.2 27.8
−2.3 ± 31.6 31.7
SKF
Wavelet
2.6 ± 15.2 15.4
−6.4 ± 20 21
Pon
12.4 ± 13.6 18.4
6.6 ± 10.2 12
QRSon
1.4 ± 3.6 3.8
0.01 ± 0.1 0.1
Rpeak
1.9 ± 13.8 13.9
−5.7 ± 8.5 10.3
QRSoff
7.5 ± 27.5 28.5
−0.01 ± 0.4 0.4
Tpeak
7.3 ± 32.2 33
0.6 ± 10.8 10.8
Toff
Table 8. Mean ± standard deviation (first line) and RMSE (second line) of error in ms between FPs and manual annotations for signals of the QT database (fs = 250Hz).
212 E. Kandi´c and L. G. Pokvi´c
Filters for Electrocardiogram Signal Processing: A Review
213
Fig. 2. Block diagram of Kalman filter.
Fig. 3. Block diagram of Wiener filter.
specified parameters, making the algorithm easier to implement. ANF is better than conventional notch filters because ANF does not only reduce unwanted effects but also preserves QRS-complex features in the filtered signal. The compared results found that the ANF has the smallest maximal value and RMS value of construction errors among the three methods, indicating an improved SNR in the filtered signal (Table 10).
4.4499
Kalman
Composite noise
Baseline wander
Muscle arti fact
Power line interference
5.3435
Weiner
Gaussian
Average
6.4427
4.4744
Kalman
3.3259
Weiner
4.8079
Kalman
5.3188
Kalman
Weiner
6.9001
4.7660
Kalman
Weiner
6.9821
Weiner
5.1929
4.9244
3.7172
4.8212
5.2970
5.3253
5.4244
6.4314
5.2543
5.9737
6.3193
7.3395
4.4944
5.8269
7.4177
8.9682
6.4117
9.2382
6.7547
8.2075
5.3288
6.2355
3.8458
5.152
6.0111
7.0645
5.5340
7.5505
5.4363
6.5083
0.1560
0.0799
0.2758
0.1763
0.1292
0.0683
0.1510
0.0741
0.2171
0.1343
Data
Data
Data
Data
MSE
SNR
Estimated signal
FiIters
Noises
0.1766
0.2337
0.2869
0.2270
0.1863
0.1971
0.1434
0.1255
0.1916
0.1422
Data
0.1462
0.0965
0.3083
0.1731
0.0929
0.0492
0.1242
0.0480
0.1239
0.0649
Data
0.1596
0.1367
0.2903
0.1921
0.1361
0.1048
0.1395
0.08253
0.1775
0.1138
Average
81546
60.790
107.49
82.574
74.109
56.615
80.89
57.026
89.03
74_384
Data
PRD
Table 9. Values of various parameters for different noises using Kalman and Wiener filter.
74.267
80.254
96.086
80.823
74.152
74.624
69.733
61.694
74.850
65.958
Data
62.736
52.918
86.603
67.861
52.178
40.452
60.351
39.436
58.439
45.492
Data
72.849
64.654
96.726
77.086
66.813
57.230
70.324
52.718
74.106
61.94
Average
214 E. Kandi´c and L. G. Pokvi´c
Filters for Electrocardiogram Signal Processing: A Review
215
Table 10. Comparisons between the ANFwSR and two types of IIR notch filters. The proposed method
The conventional IIR filter
The improved IIR filter
Entire signal
Ringing part
Entire signal
Ringing part
Maximal error
1.1601
9.0063
6.0187
370.58
310.58
RMSE
0.4436
1.8412
1.0015
60.7558
60.9910
4 Conclusion A review of denoising techniques has been conducted in this paper. The paper demonstrates how the filters and transformations play a crucial role in eliminating noise and enhancing the input ECG signal. Starting from the notch filter, where only one particular noise frequency (50 Hz) was removed at a time efficiently, but instantly failed when there was a variation in frequency of noise. Hence, the adaptive filter was introduced in order to overcome such drawbacks. In the end, it turned out that by looking at the research based on the last five years, wavelets are the most represented. Overall, these studies demonstrate ongoing research into improving the effectiveness of filters in ECG signal processing, with new and innovative techniques being developed to address specific challenges and improve the quality of ECG measurements.
References 1. Dai, B., Bai, W.: Denoising ECG by adaptive filter with empirical mode decomposition (2021). arXiv preprint arXiv:2108.08376 2. Prashar, N., Sood, M., Jain, S.: Design and performance analysis of cascade digital filter for ECG signal processing. (2019) 3. Khosravy, M., Gupta, N., Patel, N., Senjyu, T., Duque, C.A.: Particle swarm optimization of morphological filters for electrocardiogram baseline drift estimation. In: Dey, N., Ashour, A., Bhattacharyya, S. (eds.) Applied Nature-Inspired Computing: Algorithms and Case Studies. Springer Tracts in Nature-Inspired Computing. Springer, Singapore (2020). https://doi.org/ 10.1007/978-981-13-9263-4_1 4. Huang, H., Hu, S., Sun, Y.: a discrete curvature estimation based low-distortion adaptive Savitzky–Golay filter for ECG denoising. Sensors (Basel). 19(7), 1617 (2019). https://doi. org/10.3390/s19071617. PMID: 30987283; PMCID: PMC6479804 5. de Oliveira, B.R., et al.: A wavelet-based method for power-line interference removal in ECG signals. Res. Biomed. Eng. 34, 73–86 (2018) 6. Li, W.: Wavelets for electrocardiogram: overview and taxonomy. IEEE Access 7, 25627– 25649 (2018) 7. Tihak, A., Konjicija, S., Boskovic, D.: Deep learning models for atrial fibrillation detection: a review. In: 2022 30th Telecommunications Forum (TELFOR). IEEE (2022) 8. Malghan, P.G., Hota, M.K.: A review on ECG filtering techniques for rhythm analysis. Res. Biomed. Eng. 36, 171–186 (2020) 9. Venkatesan, C., Karthigaikumar, P., Paul, A., Satheeskumaran, S., Kumar, R.: ECG signal preprocessing and SVM classifier-based abnormality detection in remote healthcare applications. IEEE Access. 6, 9767–9773 (2018)
216
E. Kandi´c and L. G. Pokvi´c
10. Manju, B.R., Sneha, M.R.: ECG denoising using wiener filter and Kalman filter. Procedia Comput. Sci. 171, 273–281 (2020) 11. Chen, B., et al.: Removal of power line interference from ECG signals using adaptive notch filters of sharp resolution. IEEE Access 7, 150667–150676 (2019) 12. Petráš, I.: Novel generalized low-pass filter with adjustable parameters of exponential-type forgetting and its application to ECG signal. Sensors 22(22), 8740 (2022) 13. Akhbari, M., et al.: ECG fiducial point extraction using switching Kalman filter. Comput. Methods Programs Biomed. 157, 129–136 (2018) 14. Mihov, G.S., Badarov, D.H.: Application of a reduced band-pass filter in the extraction of power-line interference from ECG signals. In: 2020 XXIX International Scientific Conference Electronics (ET). IEEE (2020) 15. Enderle, J., Bronzino, J. (eds.) Introduction to Biomedical Engineering. Academic Press (2012) 16. Pan, J., Tompkins, W.J.: A real-time QRS detection algorithm. IEEE Trans. Biomed. Eng. 32(3), 230–236 (1985). https://doi.org/10.1109/TBME.1985.325532. PMID: 3997178 17. Hamilton, P.S., Tompkins, W.J.: Quantitative investigation of QRS detection rules using the MIT/BIH arrhythmia database. IEEE Trans. Biomed. Eng. BME-33, 1157–1165 (1986) 18. Castells-Rufas, D., Carrabina, J.: Simple real-time QRS detector with the MaMeMi filter. Biomed. Signal Process. Control 21, 137–145 (2015) 19. Yochuma, M., Renaudb, C., Jacquira, S.: Automatic detection of P, QRS and T patterns in 12 leads ECG signal based on CWT. Biomed. Signal Process. Control 25, 46–52 (2016) 20. Benitez, D.S., Gaydecki, P.A., Zaidi, A., Fitzpatrick, A.P.: A new QRS detection algorithm based on the Hilbert transform. In: Proceedings of the Computers in Cardiology 2000, Cambridge, MA, USA, 24–27 September 2000, pp. 379–382 21. Christov, I.I.: Real time electrocardiogram QRS detection using combined adaptive threshold. Biomed. Eng. Online 3, 28 (2004) 22. Zhang, F., Lian, Y.: Effective ECG QRS detection based on multiscale mathematical morphology filtering. IEEE Trans. Biomed. Circuits Syst. 3, 220–228 (2009) 23. Kim, J., Shin, H.: Simple and robust realtime QRS detection algorithm based on spatiotemporal characteristic of the QRS complex. PLoS One 11, e0150144 (2016) 24. Martinez, J.P., Almeida, R., Olmos, S., Rocha, A.P., Laguna, P.: A wavelet-based ECG delineator: evaluation on standard databases. IEEE Trans. Biomed. Eng. 51(4), 570–581 (2004) 25. Lin, C., Mailhes, C., Tourneret, J.Y.: P- and T-wave delineation in ECG signals using a Bayesian approach and a partially collapsed Gibbs sampler. IEEE Trans. Biomed. Eng. 57(12), 2840–2849 (2010) 26. Sayadi, O., Shamsollahi, M.B.: A model-based Bayesian framework for ECG beat segmentation. Physiol. Meas. 30, 335–352 (2009) 27. He, H., Tan, Y.: A novel adaptive wavelet thresholding with identical correlation shrinkage function for ECG noise removal. Chin. J. Electron. 27(3), 507–513 (2018) 28. Alyasseri, Z.A.A., Khader, A.T., Al-Betar, M.A., Abualigah, L.M.: ECG signal denoising using β-hill climbing algorithm and wavelet transform. In: Proceedings of the 8th International Conference on Information Technology, Amman, Jordan, May 2017, pp. 96–101 29. Hesar, H.D., Mohebbi, M.: An adaptive particle weighting strategy for ECG denoising using marginalized particle extended Kalman filter: an evaluation in arrhythmia contexts. IEEE J. Biomed. Health Inf. 21(6), 1581–1592 (2017)
Filters for Electrocardiogram Signal Processing: A Review
217
30. Singh, O., Sunkaria, R.K.: ECG signal denoising via empirical wavelet transform. Australas. Phys. Eng. Med. 40(1), 219–229 (2017) 31. Malik, S.A., Parah, S.A., Bhat, G.M.: Electrocardiogram (ECG) denoising method utilizing Empirical Mode Decomposition (EMD) with SWT and a Mean based filter. In: 2021 2nd International Conference on Intelligent Engineering and Management (ICIEM), London, United Kingdom, 2021, pp. 322–326. https://doi.org/10.1109/ICIEM51511.2021.9445297
Medical Physics, Biomedical Imaging and Radiation Protection
Design, Manufacturing and Quality Assessment of 3D-Printed Anthropomorphic Breast Phantom for Mammography Elma Huselji´c1 , Senad Odžak1 , Adnan Beganovi´c1,2(B) , Almasa Odžak1 , Adi Pandži´c3 , and Merim Jusufbegovi´c2 1 Faculty of Science, University of Sarajevo, Zmaja od Bosne 33-35, 71000 Sarajevo, Bosnia
and Herzegovina {senad.odzak,adnan.beganovic,almasa.odzak}@pmf.unsa.ba 2 Clinical Center of the University of Sarajevo, Bolniˇcka 25, 71000 Sarajevo, Bosnia and Herzegovina [email protected] 3 Faculty of Mechanical Engineering, University of Sarajevo, Vilsonovo šetalište 9, 71000 Sarajevo, Bosnia and Herzegovina
Abstract. Radiological anthropomorphic phantoms are objects that mimic patient’s anatomy when imaged by x-rays. These objects play a major role in optimizing radiation dose and image quality, allowing repeated exposure without exposing patients to harmful ionizing radiation. The goal of this study was to create a three-dimensionally-printed (3D-printed) phantom that would allow production of images that closely match those of human breast produced in digital mammography. We determined the attenuation properties of the Gray V4 resin using an imaged step-wedge. Gray values in image were associated with corresponding material thickness and a real mammogram was converted to two-dimensional (2D) matrix with elements whose values correspond to material thickness. Geometrical corrections for perspective and projection were taken into account. A standard triangle language (STL) file was used as input for the 3D printer. The quality of the printed phantom was evaluated by comparing its images to those of the real patient mammogram using different quantifying measures in spatial and frequency domain. The calculated similarity index (SSIM) was approximately 0.99, which indicates that SLA 3D printing technology can be successfully utilized to produce mammography phantoms. Keywords: 3D printing · Mammography · Phantom · Image similarity
1 Introduction Mammography is the most commonly used imaging modality in clinical practice for screening women for breast abnormalities and is a key tool for the early detection of breast cancer. Physical phantoms are commonly used as surrogates for breast tissue to evaluate the performance of mammography systems [1–4]. Recent advances in technology, such © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 221–231, 2024. https://doi.org/10.1007/978-3-031-49062-0_24
222
E. Huselji´c et al.
as 3D printing, have enabled the development of complex, anatomically realistic breast phantoms that have the potential to improve the evaluation of mammography systems [5]. In this study, we developed a reproducible method for creating breast phantoms using 3D printing technology. To achieve this, we developed custom software in MATLAB that takes a standard DICOM 2D mammography image as input and produces a 3D triangle mesh object as output that represents the observed attenuation of the original breast. The generated triangle mesh is stored in a generic format, which can be easily converted to the popular standard triangulation or standard tessellation language format (STL) used by most printers. The gray level of each pixel in the mammography image provides information about the radiological thickness of the breast in the direction from the focal point of the x-ray source to the pixel [6]. By taking into account the differences in x-ray attenuation between the breast tissue and 3D printing material, the observed radiological thickness in pixels can be converted to the thickness of the printed material that will produce an equivalent amount of x-ray attenuation. This process allows for the creation of an anatomically realistic breast phantom that accurately represents the x-ray attenuation properties of actual breast tissue [7]. After the realization of the phantom, its success was tested, analyzing the quality of the original clinical image and the created phantom image. Image quality is often characterized by well-known measures and criteria, usually related to radiometry, such as contrast, brightness, noise variance, and radiometric resolution. Image sharpness is a key indicator when measuring image quality [8].
2 Material and Methods 2.1 Clinical Mammography System Full-field digital mammography (FFDM) is an advanced imaging technology that offers high accuracy and precision in detecting breast cancer. Compared to traditional film mammography, FFDM provides greater clarity and detail in breast images, making it easier for radiologists to identify potential abnormalities. The use of FFDM has been shown to reduce the need for follow-up diagnostic procedures, as well as the number of false positives and unnecessary biopsies. In FFDM, pixel values represent the amount of x-ray radiation that has been absorbed by the breast tissue at each point in the image. These values are measured in units of grayscale and are used to create a detailed image of the breast. Areas of the breast that absorb more x-rays, such as denser tissues like tumors, will appear as brighter or whiter on the image, while areas that absorb fewer x-rays, such as fatty tissue, will appear darker or blacker on the image. DICOM (Digital Imaging and Communications in Medicine) is a standard format for medical images that include both the image data and the associated metadata. DICOM images can be classified into two main categories: “for processing” (raw) and “for presentation.” Raw images are acquired from the imaging equipment before any processing or analysis is performed. These images contain all the information captured by the imaging equipment and are typically stored in a lossless format to preserve image quality. In raw DICOM images, the pixel value is directly correlated to the air kerma received by
Design, Manufacturing and Quality Assessment of 3D-Printed
223
the detector (K i,d ) during the imaging process. The relationship between the pixel value and the dose of radiation is usually linear and can be used to calculate the radiation dose at any point in the image using a calibration curve. In mammography, the x-ray source is typically located above the breast, while the detector is located below. The breast is compressed between these two plates to ensure uniform thickness and reduce motion artifacts. The x-ray beam is collimated to minimize scatter radiation and to ensure that only the breast tissue is imaged. The x-ray beam used in mammography is typically a low-energy beam, which allows for better visualization of the breast tissue. The x-rays are absorbed differently by different types of breast tissue, such as dense or fatty tissue, allowing for the detection of abnormalities. The sourceto-detector distance ( f ), is an important parameter in medical imaging that affects the quality of the image obtained. In general, f refers to the distance between the x-ray source (focal spot) and the detector used to capture the x-ray image. In this study, Hologic Selenia Dimensions digital mammography system was used to obtain craniocaudal (CC) images during patient and phantom imaging. The patient had undergone a regular examination using the following exposure parameters: tube voltage (U) of 29.0 kV and tube current-exposure time product (Q) of 126.0 mA s, both determined by automatic exposure control (AEC). The associated compression force was 58.4 N. The phantom, however, was exposed using U = 29.0 kV and Q = 120 mA s, selected manually, with no compression. The source-to-detector distance ( f ) is the same in both cases. 2.2 Characterization of the Material Used in 3D Printing The surface of the 3D-printed phantom is defined by the material and the printing technique [9]. In this work, a phantom sample was printed using a stereolithography (SLA) printer that uses stereolithography technology and prints using liquid resin polymerized with an ultraviolet laser. The breast phantom was 3D printed with Grey V4 (Formlabs, Somerville, MA, USA), a material most commonly used for high-resolution rapid prototyping, product development, and design. Grey V4 is intended for the SLA printing of models with precise details; it has a matte finish and opaque appearance [10]. The surface of this material is very smooth. It is rigid, with medium mechanical resistance. Design freedom is limited due to the structure needed to support the model during printing. Unlike transparent resin, which is also used in SLA printing, the gray resin is excellent at showing details [11]. 2.3 Phantom Design The design of the phantom relied on a real CC mammogram with visible adipose and glandular tissue, with regions containing microcalcifications and other commonly visible objects. Their visibility relies on differences in x-ray attenuation caused by different density and atomic composition. This attenuation pattern can be simulated using an object made of a homogeneous material that changes its thickness accordingly. This process consists of three steps. The first step consisted of determining the attenuation properties of Grey V4 resin. A step-wedge was printed and imaged using the mammography unit. Gray values in
224
E. Huselji´c et al.
the raw and unprocessed mammography image were associated with the corresponding material thickness. Data was fitted using the following regression curve (Fig. 1): f (z) =
1
z−c a
b
,
(1)
where the fitting parameters have values of a = 290.1 ± 4.5, b = −0.149 ± 0.014, c = −72.2 ± 8.5 (R2 = 0.999). Figure 1 demonstrates small differences between values of a fitted curve (blue curve) and measured data (red points). This is also visible in the region of thicker material, which is shown in a zoomed-in view of this area. f (z) Rh 29 kV
Pixel value
10000
Pixel value
8000
6000
500 300 100
40
4000
50
60
Resin thickness (mm)
70
2000
0
0
10
20
30
40
Resin thickness (mm)
50
60
70
Fig. 1. Thickness of Grey V4 resin corresponding to different pixel values in raw full-field digital 1 b , (a = 290.1 ± 4.5, b = − mammography image. The regression curve is in form of p = z−c a 2 0.149 ± 0.014, c = −72.2 ± 8.5, R = 0.999).
In the second step, a real mammogram of an average-sized breast was converted to a two-dimensional (2D) pixel value matrix. The pixel values in the raw image were given values of corresponding material thickness, using the inverse function of (1): f −1 (p) = apb + c,
(2)
where p represents the pixel value. In the third step, in order to achieve the best results, some de-noising and image resolution reduction had to take place. Here, the most important step is to take into account geometrical corrections for perspective and projection, as shown in Fig. 2. X-ray image of the phantom was obtained at the distance of f = 655 mm between the focus and breast support plate. Due to differences in the material thickness different attenuation will occur along the path of x-ray beam. However, one must take into account the divergence of the beam paths in order to achieve the exact correspondence of attenuation between real breast and 3D-printed phantom. The attenuation path lengths (z) in the phantom are calculated for each pixel of the image, assuming a parallel geometric projection. The
Design, Manufacturing and Quality Assessment of 3D-Printed
225
tube housing
focus α f
compression plate breast
z r
breast support plate
Fig. 2. Schematic representation of the imaging system with relevant geometrical parameters
corrections for divergence due to perspective was calculated using z = 0 and z = f as coordinates of the breast support plate and focal point, respectively. To achieve this, following transformation needs to be used [12]: ⎤ ⎡ ⎤ M (z)x x ⎣ y ⎦ = ⎣ M (z)y ⎦ z cos(α)z ⎡
(3)
with r z , r= M (z) = 1 − , sin(α) =
2 f r +f2
x2 + y2 .
(4)
After the perspective transformation of the phantom, the surface was triangulated to produce an STL file for printing. 2.4 Testing the Performance of the Phantom We evaluated the successful performance of the phantom by utilizing an indirect method, which involved comparing the clinical images to those produced using the phantom. We used three different measures to evaluate the quality of the reconstructed phantom medical image: peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and Pearson correlation coefficient (PCC). PSNR is a widely-used metric that measures the ratio of the maximum possible power of a signal to the power of corrupting noise [13, 14]. SSIM is a commonly used metric that assesses the visual quality of an image by comparing its structure to that of a reference image [15]. PCC is a statistical measure
226
E. Huselji´c et al.
that evaluates the linear relationship between two variables [16]. Also, we used the modulation transfer function (MTF) and noise power spectrum (NPS) to evaluate the overall quality of clinical and phantom images. MTF and NPS are used to characterize the sharpness/resolution and noise of the imaging system. We obtained two medical images, the first one by imaging the patient and the second one by imaging a 3D phantom created from the original image. Using ImageJ software and the same calcification, we extracted data across the line of length 10 mm from both images with a pixel spacing of 0.07 mm which corresponds to detector size of 0.065 mm. We used a Gaussian normal distribution as a line spread function (LSF) to fit the extracted data (x − μ)2 K 1 , (5) LSF(x) = √ exp − 2 σ 2π 2σ 2 where K is a scaling factor. From the parameters specified in the previous equation, we calculated the full width √ at half maximum (FWHM) of the LSF using the well-known formula FWHM = 2σ 2 log 2 [17]. The relation between the LSF and MTF highlights the importance of understanding the system response for accurately characterizing the resolution of medical images [18, 19]. MTF can be determined from the LSF by using the absolute value of its Fourier transform. The analytical form of the MTF is 2 2 f σ K 1 , (6) MTF(f ) = √ exp − 2 2π 2 where f is the spatial frequency. Normalized function MTF0 ( f ) associated to MTF( f ) such that MTF0 (0) = 1 can be obtained by scaling. By Eq. (6) for a given value of MTF0 we are able to determine corresponding spatial frequency. For the given m ∈ (0, 1) from the equation MTF0 ( f m ) = m, follows that √ −2 log m . (7) fm = σ Another quantity used to characterize the images was the noise power spectrum (NPS), which is a widely used method for quantifying noise characteristics in digital imaging systems. The NPS analysis provided us with an estimate of the magnitude of noise at different spatial frequencies [20]. The formula used for the calculation is NPS(fx , fy ) =
N
2 dxdy 1
DF2 (Ik (x, y) − I k ) , N Nx Ny
(8)
k=1
where DF2 is 2D discrete Fourier transform, N is the number of regions of interest (ROIs), N x and N y are the number of pixels, and dx and dy are the pixel spacing along the x and y axis, respectively. In addition, I k (x, y) and I k are the pixel intensities and average pixel intensity for the kth ROI, respectively [21].
Design, Manufacturing and Quality Assessment of 3D-Printed
227
From 2D NPS, 1D NPS can be calculated. Firstly, radial NPS is calculated by averaging over all spatial frequencies f x and f y with fr = fx2 + fy2 . (9) The range for both f x and f y is [−f N , f N ], where f N = 7.664 mm−1 is the Nyquist frequency of pixel sampling. The final 1D NPS is the logarithm of obtained radial NPS. This curve represents the distribution of noise power as a function of the radial frequency f r , and can be used to evaluate the noise characteristics of the imaging system.
3 Results and Discussion A sample of the raw CC mammographic image was converted into a mesh object using MATLAB software. The grid object is printed using an SLA printer and the final phantom was imaged in an x-ray imaging system. A qualitative comparison of the original and final image shows that the process successfully transferred the anatomical information of the breast. The two largest microcalcifications are clearly seen in the phantom image, and the contrast is similar to the contrast seen in the original image (Fig. 3). These microcalcifications can be readily seen in the printed object as tall columns extending above the shorter background columns. The results show that SLA technology can preserve very fine details on the phantom. Different methods were used to quantify the quality of the clinical and phantom images. Table 1 shows the values of metrics (PSNR, SSIM, PCC) for the original phantom image (image 0) and cropped phantom image areas (images 1–3) compared to the originals. The PSNR values for the cropped images are higher than the PSNR value for the original image. The SSIM values for all images are close to 1. Finally, the PCC values show a strong linear relationship between the original image and the phantom image. We obtained slightly higher values of PCC for 1 and 3 than for image 2. These results suggest that the cropped images are very similar to the original images, as indicated by the high SSIM values. This high-level similarity is almost the same for different pixel ranges of the images. The higher PSNR values for the cropped images further support this conclusion. The strong linear relationship between the original medical image and the cropped images, as indicated by the high PCC values, suggests that the phantom image is a good approximation of the original image. Also, we obtained the parameters of the LSF were for the patient’s original image (σ = 0.558555, K = 350.152, μ = 5.12341) and for the phantom image (σ = 0.56475, K = 377.438, μ = 5.12638) using the line over the microcalcification, as shown in Fig. 4A. For the corresponding values of the standard deviation, we obtained FWHM values of 1.330 mm and 1.315 mm for the phantom and patient images, respectively. The slightly smaller FWHM for the patient image indicates a better resolution compared to the phantom image. For the commonly used value for medical images m = 0.1 and calculated values of σ, we obtained resolutions 3.800 mm−1 and 3.842 mm−1 for the phantom and patient images, respectively. This is clearly indicated in Fig. 5. Overall, our results indicate that
228
E. Huselji´c et al.
Fig. 3. Process of phantom manufacture: (a) Raw DICOM mammography image, (a) generated STL file, (c) 3D-printed phantom, (d) Raw DICOM phantom image Table 1. Evaluation metrics for comparing the phantom image with the original patient image for different sizes Image
PSNR
SSIM
PCC
x range
y range
0
46.2495
0.9893
0.9598
[0, 1500]
[0, 1500]
1
59.5757
0.9948
0.8990
[0, 1000]
[200, 1200]
2
59.4286
0.9947
0.8502
[0, 800]
[100, 1400]
3
59.6697
0.9949
0.8955
[100, 1000]
[100, 1300]
Design, Manufacturing and Quality Assessment of 3D-Printed
229
the patient image has a better resolution and higher spatial frequency compared to the phantom image. We calculated the NPS using the analytical formula (8). Figure 4B indicates N = 100 ROIs, with N x × N y = 50 × 50. The chosen ROIs were relatively homogeneous, with no microcalcifications or high-density tissues. Figure 6 shows the logarithm of the NPS( f r ) plotted against radial frequency f r . The results were analyzed to characterize the noise properties of the images.
B
A
Fig. 4. Regions of interest used for evaluation of: (A) Modulation transfer function at a microcalcification, (B) Noise power spectrum in a relatively homogeneous part of the image.
Notably, the NPS curves for both the patient and phantom images are very similar. Some differences between two NPS curves can be observed at higher frequencies. They are caused by differences in quantum noise due to different detector air kerma K i,d and could be corrected with the introduction of a scaling factor. However, the qualitative features of both curves (general shape and local maxima) are very similar, which indicates the same structural noise, which indicates a good overall similarity between two evaluated images.
4 Conclusions The SLA 3D printing technology can be successfully utilized to produce mammography phantoms. The quality of the printed phantom was evaluated by comparing its images to those of the real patient mammogram using different quantifying measures in the spatial and frequency domain. The calculated SSIM was approximately 0.99, PSNR above 45, and PCC above 0.85. The resolution of the two images was similar, with values of 3.800 mm−1 and 3.842 mm−1 for the phantom and patient images, respectively. NPS curves for both the patient and phantom images are very similar, indicating the same structural noise. Some compromises had to be made in the reproduction of very
230
E. Huselji´c et al.
small objects due to the limiting resolution of the available SLA printers. The methods described in this study allow the manufacturing of anthropomorphic phantoms for only a fraction of the cost of similar commercially available phantoms, as well as the production of patient-specific phantoms that could be used for different purposes in medicine. Original Phantom
1.0
MTF
0.8
0.6
0.4
0.2
0.0
0
1
2
3
4
5
Spatial frequency (mm−1 )
6
7
Fig. 5. MTF curve for patient’s and phantom’s image. The dashed lines indicate values of spatial frequency at MTF = 0.1 for both images. N = 100 ROIs, Nx = 50, Ny = 50 patient phantom 1
log(NPS)
10
100
0
2
4
6
fr (mm−1 )
8
10
Fig. 6. The 1D NPS plot presented in logarithmic scale for original and phantom images with resolution 50 × 50 pixels and 100 ROIs as shown in Fig. 4.
Design, Manufacturing and Quality Assessment of 3D-Printed
231
References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21.
Bosmans, H., et al.: Radiat. Prot. Dosimetry 117(1–3), 120 (2005) Bouwman, R.W., et al.: Med. Phys. 44(11), 5726 (2017) Cockmartin, L., et al.: Phys. Med. Biol. 62(3), 758 (2017) Balta, C., et al.: Med. Phys. 45(2), 655 (2018) Schopphoven, S., Cavael, P., Bock, K., Fiebich, M., Mäder, U.: Phys. Med. Biol. 64(21), 215005 (2019) Markey, M.K.: Physics of Mammographic Imaging. CRC Press (2012) DeWerd, L.A., Kissick, M.: The Phantoms of Medical and Health Physics. Springer (2014) Martin, C., Sharp, P., Sutton, D.: Appl. Radiat. Isot. 50(1), 21 (1999) Badal, A., Clark, M., Ghammraoui, B.: J. Med. Imaging 5(3), 033501 (2018) Pandzic, A.: Group (Fig. 1) 4(5), 6 (2021) Pizzolato, N., et al.: J. Phys. Conf. Ser. 1512, 012038 (IOP Publishing, 2020) Irnstorfer, N., Unger, E., Hojreh, A., Homolka, P.: Sci. Rep. 9(1), 14357 (2019) Szeliski, R.: Computer Vision: Algorithms and Applications. Springer Nature (2022) Hore, A., Ziou, D.: In: 2010 20th International Conference On Pattern Recognition, pp. 2366– 2369. IEEE (2010) Sampat, M.P., Wang, Z., Gupta, S., Bovik, A.C., Markey, M.K.: IEEE Trans. Image Process. 18(11), 2385 (2009) Neto, A.M., Victorino, A.C., Fantoni, I., Zampieri, D.E., Ferreira, J.V., Lima, D.A.: In: 2013 13th International Conference on Autonomous Robot Systems, pp. 1–6. IEEE (2013) Ekström, P.: Statistics and the Treatment of Experimental Data. Lunds universitet (1996) Kao, Y.H., Albert, M., Carton, A.K., Bosmans, H., Maidment, A.D.: In: Medical Imaging 2005: Physics of Medical Imaging, vol. 5745, pp. 1199–1208. SPIE (2005) Manzanares, A., Calvo, M., Chevalier, M., Lakshminarayanan, V.: Appl. Opt. 36(19), 4362 (1997) Park, H.S., Kim, H.J., Cho, H.M., Jung, J., Lee, C.L.: In: 2008 IEEE Nuclear Science Symposium Conference Record, pp. 4378–4383. IEEE (2008) Dolly, S., Chen, H.C., Anastasio, M., Mutic, S., Li, H.: J. Appl. Clin. Med. Phys. 17(3), 392 (2016)
Voxelization: Multi-target Optimization for Biomedical Volume Rendering Elena Denisova1,3,4(B) , Leonardo Manetti3,4 , Leonardo Bocchi1,3 , and Ernesto Iadanza2 1 Department of Information Engineering, University of Florence, Via S. Marta 3, 50139
Florence, Italy [email protected] 2 Department of Medical Biotechnologies, University of Siena, Via Banchi di Sotto 55, 53100 Siena, Italy 3 Epica Imaginalis, Via Rodolfo Morandi 13/15, 50019 Sesto Fiorentino, Italy 4 Eidolab, Florence, Italy
Abstract. Almost all existing software for visualization of biomedical volumes provides three-dimensional (3D) rendering. The most common techniques for 3D rendering of volume data are maximum intensity projection (MIP) and direct volume rendering (DVR). Recently, rendering algorithms based on Monte-Carlo path tracing (MCPT) have also been considered. Depending on the algorithm, level of detail, volume size, and transfer function, rendering can be quite slow. In this paper, we present a simple and intuitive voxelization method for biomedical volume rendering optimization. The main advantage of the proposed method, besides the fast structure construction and traversal, is its straightforward application to MIP, DVR and MCPT rendering techniques (multi-target optimization). The same single structure (voxel grid) can be used for empty space skipping, optimized maximum intensity calculation and advanced Woodcock tracking. The performance improvement results suggest the use of the proposed method especially in cases where different rendering techniques are combined. Keywords: Rendering optimization · Voxelization · Biomedical volumes visualization
1 Introduction It is well known that 3D representation of biomedical volumes obtained by computed tomography plays an essential role in faster understanding of patient anatomy, surgical planning, clinician training and communication with patients [8, 12, 13, 15]. For this reason, almost all existing biomedical imaging software offers different modalities for 3D volume rendering. These usually include maximum intensity projection (MIP) and direct volume rendering (DVR). Recently, probabilistic algorithms have also been considered. Depending on the algorithm, level of detail, volume size, and transfer function, rendering can be slow even on powerful hardware. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 232–241, 2024. https://doi.org/10.1007/978-3-031-49062-0_25
Voxelization: Multi-target Optimization for Biomedical Volume
233
There are several optimization techniques to speed up rendering. The optimization that can undoubtedly be used for any algorithm that requires volume sampling is an empty space skipping. Such structures as distance maps, k-d trees, octrees, and linear bounded volumes, greatly increase rendering performance, especially when combined with hardware-based acceleration [4–7, 17]. In the case of MIP, optimization is usually based on the choice of sampling interpolation [10, 11]: by default, nearest-neighbour sampling is used; only when the intensity of the voxel in question is larger than the current value of the maximum intensity, interpolation is enabled. Recently, neural networkbased approaches have also been applied to MIP optimization [2]. As for probabilistic algorithms, some interesting optimization approaches can be applied to Wood cock tracking [1, 14, 18]. These approaches imply a splitting of the original volume with a large density variation into subvolumes with a small local density variation. Most of the described approaches require the construction of auxiliary structures. In these cases, the algorithmic complexity and memory storage required to construct and traverse the chosen structure should be considered. If the performance overhead overshadows the performance gain or the memory consumption is high, it is questionable whether the algorithm will be used in practice. Otherwise, if the acceleration structure covers more than one optimization aspect in combination with a performance gain, it becomes an ideal challenger. In this paper, we present an innovative approach to multi-target optimization, that uses a voxel grid as a supportive structure. The idea of the acceleration structure is based on the methods proposed by [9] and [3]. We will show how it can be quickly constructed and applied to increase the performance gain of different rendering techniques. The performance results described in the final part of this paper suggest the use of the proposed method especially when different rendering techniques are combined.
2 Methods In this section, we first describe the construction of the voxel grid and then show how this auxiliary structure can be used for empty space skipping, optimized MIP computation, and advanced Woodcock tracking. We also provide pseudocode of the described algorithms to facilitate reproducibility of our results. In this paper, we refer to voxels as the values in the nodes of the original volume, and super voxels as the values in the nodes of the constructed voxel grid. All the described techniques are implemented in OpenGL. 2.1 Construction The construction of a voxel grid is called voxelization. In this process, the ad jacent voxels of the original volume are combined to form a super voxel - a 3D cube of a certain size. The set of super voxels forms a voxel grid (see Fig. 1). The size of the voxel grid depends, of course, on the size of the super voxel and is equal to (X/s, Y/s, Z/s), where (X, Y, Z) is the size of the original volume and s is the size of the super voxel.
234
E. Denisova et al.
Fig. 1. Voxelization. Left: original volume; right: voxel grid for super voxel size s = 16
For our purposes, we construct a super voxel by finding the maximum value of transparency (or density) of all voxels that fall within the super voxel. Note, that whenever the transfer function is changed, the voxel grid must be reconstructed. Therefore, it is important to construct the voxel grid quickly. Our solution is to delegate the construction of the voxel grid to the OpenGL fragment shader stage. First, we set the dimensions of the viewport to the size of the volume slice X×Y. Then we draw the quad that covers the entire viewport. We repeat the drawing Z times, where Z is the number of slices in the volume, by scrolling from 0 to (Z − 1) to enclose the entire volume. To construct the voxel grid, we use the OpenGL function imageAtomicMax, which overwrites the stored value if it is smaller than the specified value (see Algorithm 1). The voxel density is calculated as follows: first, trilinear interpolation is applied to the samples of the original volume to obtain the voxel intensity; then, the voxel intensity is mapped to the current transfer function. Given the maximum volume dimension N = max(X, Y, Z) and the super voxel size s, the complexity of the voxelization process is O(N 3 ), the memory consumption is N 3 . O s 2.2 Empty Space Skipping An empty space skipping optimization is based on the idea that the voxels whose density is zero (or lower than a certain threshold) do not normally contribute to the result of a volume rendering procedure and therefore should not be sampled. Starting from a ray traversing the volume in the direction rayDir from the intersection point rayStart to the exit point rayEnd, naive volume sampling along the ray can be achieved by advancing with a fixed small step d.
Voxelization: Multi-target Optimization for Biomedical Volume
235
Algorithm 1 Voxelization viewport ← (X, Y ) s ← SuperV oxelSize i←0 glUniform3f(dimsLocation, X, Y, Z) glUniform1i(sLocation, s) while i = Z do glUniform1i(iLocation, i) glDrawElements(...) i←i+1 end while volumeCoords ← (gl F ragCoords.xy, i)/(X, Y, Z) density ← Density(volumeCoords) voxelGridCoords ← (gl F ragCoords.xy, i)/(s, s, s) imageAtomicMax(V oxelGrid, voxelGridCoords, density)
C++
GLSL
By construction, the voxel grid introduced above can be used for empty space skipping in a straightforward manner. Suffice it to note that if the density of the super voxel is less than a certain threshold, all the voxels that contributed to the construction of this super voxel in the voxelization phase can be skipped without sampling. Thus, instead of advancing with a small step d, we advance with a step D = d·s, where s is the size of the super voxel. The idea can be formalized by Algorithm 2. To advance safely with step D, the super voxel density should be checked one D step ahead. Otherwise, the boundary voxels may be erroneously skipped (see Fig. 2). Algorithm 2 Empty Space Skipping rayLen ← length(rayEnd − rayStart) threshold ← 10−5 pos ← rayStart while rayLen ≥ 0 do α ← max(SuperDensity(pos), SuperDensity(pos + D · rayDir)) step ← d if α > threshold then Sample(pos) else step ← D end if pos ← pos + step · rayDir rayLen ← rayLen − step end while
GLSL
2.3 Maximum Intensity Projection The idea of optimising maximum intensity calculation is very similar to empty space skipping. The main difference is that the density of the super voxel should be compared to the current maximum intensity value, not to a constant threshold (in the case of MIP,
236
E. Denisova et al.
Fig. 2. Empty space skipping: erroneous skip. Empty space skipping without ahead verification causes erroneous skip (see red arrows).
density and intensity are synonyms). Another advantage over Algorithm 2 is that even if the density of the super voxel is less than the current maximum intensity, and the sampling is necessary, the super voxel still can be skipped if the result of the sampling is equal to the total density of the super voxel (see Algorithm 3). 2.4 Woodcock Tracking Woodcock Tracking is the sampling technique proposed by Woodcock in 1965 [16]. Unlike the Riemman sum, it is unbiased and does not require the sampling of every point along the traversing ray. Instead of choosing the transmittance value, the distance at which this value could be obtained is chosen as a function of the maximum density α max in the entire volume. The voxel at the probe distance is accepted with probability α αmax , where α is the density of the probe voxel. If the voxel is rejected, the process is repeated from the rejected point until a voxel is accepted or the end of the ray is reached. However, if the local volume density is much smaller than the maximum volume density, the gain of Woodcock Tracking smooths out (the probability of accepting voxels from low density regions becomes very low, and the process should be repeated until the high-density region is reached). To solve this problem, an advanced Woodcock Tracking has been proposed [1, 14, 18]. It suggests to divide the original volume with large density variation into subvolumes with small local density variation. The probe distance is then chosen as a function of α , where the maximum local density, and the probability of accepting a voxel is α(pos) max α(pos)max is the density of the super voxel corresponding to the probe voxel.
Voxelization: Multi-target Optimization for Biomedical Volume
237
Algorithm 3 Maximum Intensity Calculation rayLen ← length(rayEnd − rayStart) αmax ← 0 threshold ← 1 − 10−5 pos ← rayStart while rayLen ≥ 0 do α ← max(SuperDensity(pos), SuperDensity(pos + D · rayDir)) if α > αmax then i ← Density(pos) αmax ← max(αmax , i) if αmax > threshold then break end if end if step ← d if αmax ≥ α then step ← D end if pos ← pos + step · rayDir rayLen ← rayLen − step end while
GLSL
It is easy to see, that the previously constructed voxel grid can be used to implement advanced Woodcock Tracking (see Algorithm 4). Note that if the super voxel has a very small maximumdensity, the distance l generated on this super voxel will be very large log(rand ) . Therefore the distance should be adjusted based on the maximum l = − α(pos) max density alphamax , otherwise there is a risk of moving too far away from the current voxel and skipping the voxels with significant density. This would lead to the situation shown in Fig. 2.
3 Results In this section, we illustrate the results of our optimization method applied to three different CT acquisitions rendered using the following techniques: DVR (empty space skipping according to Alg. 2), MIP (according to Alg. 3), and Monte-Carlo Path Tracing (MCPT) (volume sampling is implemented using advanced Woodcock Tracking as described in Alg. 4). We apply three typical trans fer functions to all acquisitions (see Fig. 3) and present speed measurements of voxelization (Table 1) and rendering (Table 2). As shown in Table 1, the voxelization process takes between ∼300 and 650 ms depending on the volume size and structure complexity. However, voxelization does not occur during interaction with the volume, but when the transfer function changes. Therefore, in practice, it does not affect interactivity.
238
E. Denisova et al.
Fig. 3. Test data sets Algorithm 4 Advanced Woodcock Tracking l ← − log(rand) αmax rayLen ← length(rayEnd − rayStart) − l αcurr ← αmax threshold ← 10−5 pos ← rayStart + l · rayDir while rayLen ≥ 0 do α ← max(SuperDensity(pos), SuperDensity(pos + D · rayDir)) l←D if α > threshold then i ← Density(pos) i > rand then if αcurr break end if l ← − log(rand) α if l ≥ D then l ← − log(rand) αmax αcurr ← αmax else αcurr ← α end if end if pos ← pos + l · rayDir rayLen ← rayLen − l end while
GLSL
All data sets are encoded in 16-bit format. Data sets: Cat 803 × 803 × 598 (1a–1d), Hand 707 × 704 × 1336 (2), Knee 704 × 703 × 1327 (3). Transfer Functions: Bone (1a), Skin (1b), Mixed (1c), Full (1d). Bone, Skin, Mixed are applied on Direct Volume Rendering (DVR) and Monte-Carlo Path Tracing (MCPT), while Full is applied only to Maximum Intensity Projection (MIP). Voxelization is measured in milliseconds. The super voxel size is 4. All performed on Intel(R) Core(TM) i5-7600K CPU @ 3.80 GHz with NVIDIA GeForce GTX 1060 4Gb. Table 2 shows that the proposed voxelization technique increases the performance of DVR up to 9 times. The speed-up strongly depends on the transfer function. The transfer functions that zero out a larger volume area result in a major performance gain (see Bone). In the case of MCPT, the 5 times acceleration is stable and does not depend on the transfer function. This is an expected situation and shows that the advanced
Voxelization: Multi-target Optimization for Biomedical Volume
239
Table 1. Voxelization speed measurement Bone
Skin
Mixed
Full
Cat
303
309
322
306
Hand
539
546
548
648
Knee
534
576
600
627
Woodcock Tracking has solved the problem of large density variations introduced by some transfer functions (see Mixed). MIP acceleration is also stable and is about 50% of the original velocity. The minor MIP speed-up can be explained by the early termination when alphamax becomes larger than threshold (see Alg. 3). Table 2. Rendering Speed Measurement Bone w Cat
DVR
Skin w/o r
w
Mixed w/o r
w
Full w/o r
21.3 3.0
7.1× 20.4 4.4
4.6×
7.4 2.8
2.6×
MCPT 13.3 3.1
4.3× 14.3 3.5
4.1×
4.1 1.2
3.4×
MIP Hand DVR
w/o
r
27.0 19.2 1.4× 20.0 2.2
9.1× 22.2 2.5
8.9× 10.4 2.1
5.0×
MCPT 12.5 2.5
5.0× 15.2 2.7
5.6×
5.3×
6.4 1.2
MIP Knee DVR
w
23.8 14.5 1.6× 21.3 2.4
8.9× 19.6 4.1
4.8×
6.5 2.1
3.1×
MCPT 14.5 2.9
5.0× 10.6 2.7
3.9×
3.5 0.8
4.4×
MIP
19.6 14.7 1.3×
The speed for Direct Volume Rendering (DVR), Monte-Carlo Path Tracing (MCPT) and Maximum Intensity Projection (MIP) is given in frames per second. Each column contains the speed measurement of the algorithm with optimization (w), without optimization (w/o), and acceleration rate (r). Super voxel size is 4. All preformed on Intel(R) Core(TM) i5-7600K CPU @ 3.80 GHz with NVIDIA GeForce GTX 1060 4Gb, viewport size 1727 × 822.
4 Discussion and Conclusion The results of our research suggest the use of the proposed method especially in cases where different rendering techniques are combined. Thanks to the proposed method, real-time interaction with data obtained by computed tomography becomes feasible even for such a greedy algorithm as Monte-Carlo path tracing.
240
E. Denisova et al.
The images presented in this paper were generated from post mortem (Hand, Knee), or animals (Cat) CBCT (Cone-Beam Computed Tomography) acquisitions obtained with the HDVI (High-Definition Volumetric Imaging) CT Medical Imaging Platform. Our further research will focus on finding the optimal size of the super voxel in relation to the complexity of the volume structure.
References 1. Behlouli, A., Visvikis, D., Bert, J.: Improved woodcock tracking on monte carlo simulations for medical applications. Phys. Med. Biol. 63(22), 225005 (2018) 2. Chao, Z., Xu, W.: A new general maximum intensity projection technology via the hybrid of u-net and radial basis function neural network. J. Digit. Imaging 34(5), 1264–1278 (2021) 3. Crassin, C., Neyret, F., Lefebvre, S., Eisemann, E.: Gigavoxels: ray-guided streaming for efficient and detailed voxel rendering. In: Proceedings of the 2009 Symposium on Interactive 3D graphics and games, pp. 15–22 (2009) 4. Crassin, C., Neyret, F., Sainz, M., Green, S., Eisemann, E.: Interactive indirect illumination using voxel cone tracing. In: Computer Graphics Forum, vol. 30, pp. 1921–1930. Wiley Online Library (2011) 5. Deakin, L.J., Knackstedt, M.A.: Efficient ray casting of volumetric images using distance maps for empty space skipping. Comput. Visual Media 6, 53–63 (2020) 6. Hadwiger, M., Al-Awami, A.K., Beyer, J., Agus, M., Pfister, H.: SparseLeap: Efficient empty space skipping for large-scale volume rendering. IEEE Trans. Visual Comput. Graph. 24(1), 974–983 (2017) 7. Ize, T., Wald, I., Parker, S.G.: Ray tracing with the bsp tree. In: 2008 IEEE Symposium on Interactive Ray Tracing, pp. 159–166. IEEE (2008) 8. Kutaish, H., Acker, A., Drittenbass, L., Stern, R., Assal, M.: Computer-assisted surgery and navigation in foot and ankle: state of the art and fields of application. EFORT Open Rev. 6(7), 531 (2021) 9. LaMar, E., Hamann, B., Joy, K.I.: Multiresolution Techniques for Interactive Texture-Based Volume Visualization. IEEE (1999) 10. Mroz, L., König, A., Gröller, E.: Real-time maximum intensity projection. In: Data Visualization’99: Proceedings of the Joint EUROGRAPHICS and IEEE TCVG Symposium on Visualization in Vienna, Austria, May 26–28, 1999, pp. 135–144. Springer (1999) 11. Mroz, L., König, A., Gröller, E.: Maximum intensity projection at warp speed. Comput. Graph. 24(3), 343–352 (2000) 12. Pachowsky, M.L., et al.: Cinematic rendering in rheumatic diseases—photorealistic depiction of pathologies improves disease understanding for patients. Front. Med. 9 (2022) 13. Sariali, E., Mauprivez, R., Khiami, F., Pascal-Mousselard, H., Catonné, Y.: Accuracy of the preoperative planning for cementless total hip arthroplasty. a randomised comparison between three-dimensional computerised planning and conventional templating. Orthopaed. Traumatol. Surg. Res. 98(2), 151–158 (2012) 14. Szirmay-Kalos, L., Tóth, B., Magdics, M.: Free path sampling in high resolution inhomogeneous participating media. In: Computer Graphics Forum, vol. 30, pp. 85–97. Wiley Online Library (2011) 15. Wang, C., et al.: Patient-specific instrument-assisted minimally invasive internal fixation of calcaneal fracture for rapid and accurate execution of a preoperative plan: a retrospective study. BMC Musculoskelet. Disord. 21, 1–11 (2020)
Voxelization: Multi-target Optimization for Biomedical Volume
241
16. Woodcock, E., Murphy, T., Hemmings, P., Longworth, S.: Techniques used in the gem code for monte carlo neutronics calculations in reactors and other systems of complex geometry. In: Proc. Conf. Applications of Computing Methods to Reactor Problems, vol. 557. Argonne National Laboratory (1965) 17. Zellmann, S.: Comparing hierarchical data structures for sparse volume rendering with empty space skipping (2019). arXiv preprint arXiv:1912.09596 18. Zhou, S.: Woodcock tracking based fast Monte Carlo direct volume rendering method. J. Syst. Simul. 29(5), 1125 (2017)
Importance of Patient Dose Evaluation and Optimization in Thorax Computed Tomography ˇ 2 , Mustafa Busuladži´c3 , Azra Gazibegovi´c-Busuladži´c4 , Belkisa Hani´c1 , Lejla M. Civa Amra Skopljak-Beganovi´c5 , and Adnan Beganovi´c4,5(B) 1 General Hospital “Prim. Dr. Abdulah Nakˇas” Sarajevo, 71000 Sarajevo, Bosnia and
Herzegovina [email protected] 2 University “Sarajevo School of Science and Technology”, Sarajevo Medical School, 71210 Ilidža, Bosnia and Herzegovina [email protected] 3 Faculty of Medicine, University of Sarajevo, 71000 Sarajevo, Bosnia and Herzegovina [email protected] 4 Faculty of Science, University of Sarajevo, 71000 Sarajevo, Bosnia and Herzegovina {azragb,adnan.beganovic}@pmf.unsa.ba, [email protected] 5 Department of Radiation Protection and Medical Physics, Clinical Center of the University of Sarajevo, 71000 Sarajevo, Bosnia and Herzegovina [email protected]
Abstract. Computed tomography (CT) is one of the most important imaging modalities in modern medicine. Using CT, one can obtain useful information about a patient’s health status, or condition. As this modality delivers some amount of radiation to the patient’s body that can be harmful, special attention must be paid to choosing appropriate parameters that can reduce the dose but maintain helpful diagnostic information from the CT image. Radiation doses for CT examinations vary considerably among patients, institutions, and countries. This variation is mostly attributable to the technical parameters of the CT scanning protocols. An optimization process is a team effort of the CT radiologist, the lead CT technologist, and the clinically qualified medical physicist. The purpose of this study is to analyze patient doses in thorax CT in two public hospitals in Sarajevo, the capital of Bosnia and Herzegovina. Data were collected from five different CT devices from these hospitals, using the OpenREM system. The optimization process in both hospitals was surveyed as well, investigating their benefits and shortcomings. Finally, we propose possible ways for future optimization and harmonization of existing protocols in two hospitals, by adjusting different technical parameters such as the tube voltage, tube current, and pitch. Keywords: Computed tomography · Patient dose · Optimization
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 242–251, 2024. https://doi.org/10.1007/978-3-031-49062-0_26
Importance of Patient Dose Evaluation and Optimization
243
1 Introduction Thorax computed tomography (CT) is a commonly used imaging technique for the diagnosis and follow-up of various diseases affecting the lungs, mediastinum, and chest wall [1, 2]. It plays a key role in the management of trauma patients [3, 4]. CT is recognized as a method of choice in early lung cancer screening [5–9]. Like many other imaging techniques that rely on x-rays, it is associated with exposure to ionizing radiation. As technology has advanced and the cost of equipment has decreased, the use of CT has become more widespread. This should be regarded as a positive development, as CT is a valuable tool in modern medicine that provides information necessary for accurate diagnosis and treatment planning. However, the increasing use of CT scans has raised concerns about the potential risks associated with ionizing radiation, including an increased risk of cancer [10]. Patient dose evaluation and optimization is an important aspect of CT imaging that aims to minimize patient exposure to ionizing radiation while maintaining diagnostic image quality [11, 12]. Several factors, including patient size, CT scanner settings, and imaging protocols, can impact the radiation dose delivered to the patient during thorax CT scans. Therefore, it is crucial to assess the dose delivered to each patient and optimize the imaging protocol accordingly [1]. Optimization, however, is a challenging process. In general, increasing image quality and lowering the patient dose are two goals that sometimes conflict, and finding a balance that achieves both can be challenging. Furthermore, patients vary in size and anatomy characteristics. Larger patients may require higher radiation doses to produce adequate image quality. The desired imaging parameters, sometimes, may not be achieved due to the technical limitations of CT scanners. Moreover, the lack of standardization in imaging protocols, as well as lack of awareness and understanding of radiation risks, could lead to unnecessary exposures and affect the overall patient doses [13]. An increasing number of countries are implementing lung cancer screening programs that involve the use of CT [5–9]. By detecting lung cancer early, these screening programs can help to reduce the burden of disease and improve patient outcomes. However, there are also concerns about the potential harms of CT based screening, such as the risk of false positives, overdiagnosis, and radiation exposure. In such cases, the reduction of patient doses becomes an imperative, and the optimization process a necessity [9]. This study aims to explore patient doses and optimization practices in thorax CT in two public hospitals in Sarajevo, Bosnia and Herzegovina.
2 Materials and Methods 2.1 Data Collection The study evaluated the dose data of 3008 CT chest procedures performed on patients who were admitted to one of the public hospitals in Sarajevo – Clinical Centre of the University of Sarajevo (KCUS), or General Hospital “Prim. dr. Abdulah Nakˇas” Sarajevo (OBS), in a 12-month period. Dose descriptors and accompanying procedure information were collected by the OpenREM dose monitoring system (The Royal Marsden NHS Foundation Trust, London, United Kingdom) which interprets radiation-structured dose reports (RDSR) stored on the picture archiving and communication system (PACS).
244
B. Hani´c et al.
OpenREM ensured that the patients’ identities remained anonymous. This limited our ability to access additional information about the patients, such as pregnancy status, immunodeficiency, comorbidities, number of CT scans received, disease outcome, and other relevant factors. As a result, no corresponding approval from the local ethical committee was required. The patients had RDSR files available, which provided detailed information about their CT examination, including the type of examination, date and time of the procedure, patient age and sex, exposure time (t), scan length (L), slice thickness (T ), collimation width (nT ), pitch (p), tube voltage (U), maximum and mean tube current (I), rotation time, as well as the values of dose indices, specifically air kerma length product (DLP) and pitch-corrected volume CT air kerma index (CTDIvol ). It needs to be emphasized that OpenREM cannot process data that are not provided by the RDSR file itself. The study excluded patients with incomplete data. IBM’s Statistical Package for Social Sciences (SPSS) version 26.0 (International Business Machines Corporation, Armonk, New York, USA) was employed to analyze the data, using a significance level of α = 0.05 in statistical calculations. To test the normality of the distribution, the Kolmogorov-Smirnov test was utilized. Typically, the dose data are not normally distributed, so the non parametric Mann-Whitney U test was used to evaluate differences between data distributions. Patients were scanned on one of the scanners listed in Table 1. Table 1. CT scanners used in the study. ID
Manufacturer
Model
Number of slices
CT1
Toshiba
Aquilion Prime SP
160
CT2
Siemens
SOMATOM Definition AS
CT3
Toshiba
Aquilion Prime SP
CT4
Toshiba
Astelion
CT5
GE
Optima660
16 160 16 128
2.2 Evaluation of Optimization Practices Regulations in Bosnia and Herzegovina require clinically qualified medical physicist (CQMP) to be involved in practical aspects of medical exposure, such as calibration of equipment, calculation of the patient dose, development of com plex techniques, creation of quality assurance program and implementation of quality control. According to the regulations, justification and optimization aspects for examinations in diagnostic radiology are elements of a quality assurance program, whose implementation is required [14]. Another important aspect of the regulations is a requirement for full-time employment of medical physicists in hospitals that have nuclear medicine or radiotherapy departments, while departments of medical physics are required for institutions that provide all three
Importance of Patient Dose Evaluation and Optimization
245
fields that use ionizing radiation, namely, diagnostic radiology, nuclear medicine, and radiotherapy. However, the optimization process is a team effort; this team has to consist of at least the lead CT radiologist, the lead CT technologist, and CQMP. It is recommended for a senior member of the facility administration team to be involved [13]. The role of vendor application specialists should not be neglected. In cases when a new CT scanner is installed, the application specialists are the first to set up imaging protocols, which are in some cases never reviewed or examined. In this paper we surveyed the optimization process in both above-mentioned hospitals, investigating their strengths and weaknesses.
3 Results and Discussion Data collected from OpenREM dose management system allows the evaluation of various parameters related to the scanning protocol. The most important ones are summarized in Table 2 which contains information on rotation time t, pitch p, tube voltage U, average tube current I ave and maximum tube current I max for CT scanners included in the study. Unfortunately, the RDSR file does not provide all information relevant to protocol design. However, a lot can be inferred from available data. Table 2. Overview of the most important protocol parameters (rotation time t, pitch p, tube voltage U, average tube current I ave and maximum tube current I max ) for computed tomography scanners included in the study. Ns
Series 1
Series 2
t
p
U
Iave
Imax
t
p
U
Iave
Imax
CT1
2
0.35
0.813
120
137
195
0.38
0.726
120
450
500
CT2
2
0.50
1.200
120
345
650
0.50
1.200
100
334
650
CT3
1
0.50
0.813
120
127
180
CT4
1
0.75
0.980
120
71
110
CT5
1
0.70
0.980
120
400
400
The study includes patient dose and protocol data from 1549 male and 1459 female patients. Two sexes are equally distributed (binomial test, p = 0.105). Figure 1 is a histogram that shows the age and sex distribution of patients. The average age of patients is 61.9 y with interquartile range (ΔQ) of 17.7 y. The distribution is negatively skewed (s = −0.692) and not normal (Kolmogorov Smirnov test, p < 0.001). The age distribution is the same between the two sexes (Mann-Whitney U test, p = 0.963). Figures 2 and 3 show the relevant patient dose quantities, DLP and C VOL , respectively, while Table 3 shows median and interquartile range values of these quantities.
246
B. Hani´c et al. Sex
Number of patients
400
Male Female
300 200 100 0
0
20
40
60
80
100
Age Fig. 1. Number of patients relative to age/sex.
Table 3. Median (˜x) and interquartile range (ΔQ) of air kerma-length product (PKL,CT or DLP) and volume computed tomography air kerma index (C VOL or CTDIVOL ) for patients examined on CT scanners included in the study. PKL,CT (Gycm) CT scanner
CVOL (mGy)
x˜
Q
x˜
Q
CT1
640
212
7.8
9.0
CT2
413
381
6.5
3.4
CT3
335
209
7.2
4.6
CT4
316
240
7.1
5.7
CT5
1039
157
24.8
1.6
Two hospitals were surveyed on their practices in the optimization process. The positive outcome of the regulations in medical exposure is the availability of medical physicists specialized in medical imaging – 4 in KCUS and 1 in OBS. However, no teams are officially committed to reviewing and management of imaging protocols. In KCUS, a provisional team was established to follow up results of national technical cooperation projects with International Atomic Energy Agency (IAEA). This team was involved in practical aspects of optimization in CT imaging. In OBS, however, no such team exists, and its protocol management is exclusively done by application specialists from corresponding equipment vendors. Data collected by the OpenREM dose management system provides valuable data which allows analysis of scanning protocols (Table 2). It should be noted, however, that some relevant information regarding the CT scanning protocol, primarily a user-defined
Importance of Patient Dose Evaluation and Optimization
247
DLP total (Gycm)
3
2
1
0
CT1
CT2
CT3
CT4
CT5
CT scanner Fig. 2. Boxplot of total air kerma-length product (PKL,CT or DLP) for patients examined on CT scanners included in the study. Outliers and extreme values are represented with circles and asterisks, respectively.
40
CTDIvol
30 20 10 0
CT1
CT2
CT3
CT4
CT5
CT scanner Fig. 3. Boxplot of volume computed tomography air kerma index (C VOL or CTDIVOL ) for patients examined on CT scanners included in the study. Outliers and extreme values are represented with circles and asterisks, respectively.
parameter that controls the level of image noise in CT images, is not available. This parameter is known as “noise index”, “standard deviation”, or “reference mAs.” By adjusting this parameter, the operator can increase or decrease the level of image noise, which can impact the image quality and radiation dose [13]. The noise level is set to
248
B. Hani´c et al.
a specific slice thickness (T ). In practice, the noise level is the main input parameter for the tube current modulation (TCM), whose role is to change I in order to adjust the amount of radiation emitted by the X-ray tube during the scanning process based on the patient’s anatomy [15]. Although the information on the “noise index” was not available, information on maximum and mean I allowed some insight into noise level settings. It is not beyond reason to expect for the mean I to be somewhere in middle between minimum and maximum I, on average for all patients [16]. Although data provides no information on minimum I, the available protocols recommended by the American Association of Physicists in Medicine (AAPM) indicate the recommended minimum of 80 mA for Canon/Toshiba, and 100 mA for General Electric CT scanners [17]. The obvious problem that can be seen from median values presented in Table 2 is that I ave and I max in CT5 are the same, indicating the inappropriate settings of either noise index, or minimum I. Values for other examined CT scanners could be considered satisfactory. Another value that stands out in Table 2 is U for CT2. The use of lower tube voltage is a recognized method for dose reduction and image quality optimization, especially in CT pulmonary angiography [18, 19]. The visualisation of iodine contrast is more prominent when lower U is used. This method of dose reduction requires use higher values of I, which is sometimes not possible. Hence, I max for CT2 is 650 mA. When compared to the recommended AAPM protocols, the reported values of pitch, p, are low for all scanners, except for CT2 [17]. In general, p should be above 1. It is noticeable that CT1 and CT2 use two phases for standard chest CT – non-contrast (NC) phase and contrast-enhanced (CE) phase. Use of both, NC and CE series, in the same procedure is not recommended in routine practice [20]. Effectively, patient doses are increased by factor 2 when such imaging procedure is used. Boxplots in Figs. 2 and 3 indicate distribution of patient doses in evaluated CT scanners. Table 3 indicates median (˜x) and interquartile range (Q) of air kerma-length product (PKL,CT or DLP) and volume computed tomography air kerma index (C VOL or CTDIVOL ). Total DLP refers to dose during the whole examination. This includes one or two scan projection radiographs (SPR), stationary acquisitions for contrast media tracking, as well as NC and CE phases which account for the most of radiation exposure. Values of CTDIVOL refer to a single helical acquisition, either NC or CE. The highest doses are reported for CT5. Median DLP is 1039 mGy cm with interquartile range of Q = 157 mGy cm, while median CTDIVOL is 24.8 mGy (Q = 1.6 mGy). While value of CTDIVOL is below the national diagnostic reference level (DRL) which is set to 30 mGy, DLP is well above the recommended value (650 mGy cm), which indicates that optimization is necessary [14]. Values achieved on other CT scanners are within national and international recommendations. Median DLPs for CT1, CT2, CT3 and CT4 are (640, 413, 335 and 316) mGy cm. It should be noted that CT1 and CT2 include 2 series – NC and CE. DLPs could be considerably lower if the examination was reduced to the CE phase only. DLPs for CT3 and CT4 are not significantly different (Mann Whitney U test, p = 0.055). CTDIVOL values for CT1, CT2, CT3 and CT4 are between 6.5 mGy and 7.8 mGy. These values are very close to those recommended in international publications, and well below the national DRL.
Importance of Patient Dose Evaluation and Optimization
249
Overall, optimization is necessary for CT1 and CT2, mainly due to use of both NC and CE phase in general chest CT examination. This is a problem related to the information flow and use of indication-based protocols. Unfortunately, it is not uncommon for CT technologists not to be aware of indication for chest CT. Hence, they do not know whether NC is required, or even if the delayed acquisition of chest and abdomen, which is commonly requested for base line staging and follow-up of lung cancer, is necessary. In this case, a major role should be given to the leading radiologist who should provide means for this information to be provided to technologists. In the case of CT5, high values of CTDIVOL and the same values of I ave and I max , indicate incorrect use of noise index and/or minimum I in the scanning protocol. Indeed, this has been proven to be true after the protocol was reviewed by the CQMP. The noise index (NI) was set to value of 12.5 at slice thickness of 0.625 mm. The NI wouldn’t be considered high, but the reference slice thickness was too low. As a consequence, the TCM system increased I to values close to the chosen I max . The CQMP should review the protocol and make necessary adjustments together with radiologists, as radiologists should be aware of protocol changes that would affect the image quality. Thus, the reduction of dose from 24.8 mGy to something more appropriate must be done in steps (i.e. by raising NI to 15, it is expected to achieve a 31% reduction in dose, and subsequently increasing it to 20 and continuing in the same manner until the desired result is reached). The sudden reduction of image quality may cause problems, and the new protocol could be immediately rejected by the radiologists in charge. In other words, this would lead to a decreased visibility of different tissue structures of interest and thus decreased sensitivity for observation of specific pathologies. Optimization of the protocol must be related to the tissue structure we want to sample by CT. One of the possible directions of the optimization is to use an anthropomorphic chest phantom [21]. In this sense, our next step could be the production of a homemade anthropomorphic thorax phantom to improve existing protocols for an immunocompromised patient. Recently, we used a similar approach, a 3D-printed infant head phantom, to optimize paediatric scanning protocol [22].
4 Conclusion This study provided valuable insight into how dose management software can be used in the optimization process. Although the software provides limited data on protocol and image quality, its value is immense, as it makes possible frequent check-ups of protocols used in everyday practice in diagnostic radiology. The collected data from two hospitals in Sarajevo can be used to compare patient doses to the national DRLs, and also give valuable insight into how the examinations are performed. The existence of unoptimized protocols indicates that some changes need to be made in the decision process. Teams committed to CT protocol optimization should be established, utilizing all positive aspects of national regulations and international recommendations.
250
B. Hani´c et al.
References 1. Kofler, J.M., Cody, D.D., Morin, R.L.: CT protocol review and optimization. J. Am. Coll. Radiol. 11(3), 267–270 (2014) 2. Dymbe, B., Mæland, E.V., Styve, J.R., Rusandu, A.: Individualization of computed tomography protocols for suspected pulmonary embolism: a national investigation of routines. J. Int. Med. Res. 48(4), 0300060520918427 (2020) 3. Iacobellis, F., Romano, L., Rengo, A., Danzi, R., Scuderi, M.G., Brillantino, A., et al.: CT protocol optimization in trauma imaging: a review of current evidence. Curr. Radiol. Rep. 8, 1–9 (2020) 4. Flammia, F., Chiti, G., Trinci, M., Danti, G., Cozzi, D., Grassi, R., et al.: Optimization of CT protocol in polytrauma patients: an update. Eur. Rev. Med. Pharmacol. Sci. 26, 2543–2555 (2022) 5. van der Aalst, C.M., Ten Haaf, K., de Koning, H.J.: Implementation of lung cancer screening: what are the main issues? Transl. Lung Cancer Res. 10(2), 1050 (2021) 6. Copeland, A., Criswell, A., Ciupek, A., King, J.C.: Effectiveness of lung cancer screening implementation in the community setting in the United States. J. Oncol. Pract. 15(7), e607– e615 (2019) 7. Kim, Y.: Implementation of organized lung cancer screening program in Korea. Ann. Oncol. 30, ii14 (2019) 8. Rzyman, W., Szurowska, E., Adamek, M.: Implementation of lung cancer screening at the national level: Polish example. Transl. Lung Cancer Res. 8(Suppl 1), S95 (2019) 9. Vonder, M., Dorrius, M.D., Vliegenthart, R.: Latest CT technologies in lung cancer screening: protocols and radiation dose reduction. Transl. Lung Cancer Res. 10(2), 1154 (2021) ˇ 10. Civa, L.M., Beganovi´c, A., Busuladži´c, M., Jusufbegovi´c, M., Awad-Dedi´c, T., VegarZubovi´c, S.: Dose descriptors and assessment of risk of exposure-induced death in patients undergoing COVID-19 related chest computed tomography. Diagnostics 12(8), 2012 (2022) 11. Singh, S., Kalra, M.K., Khawaja, R., Padole, A., Pourjabbar, S., Lira, D., et al.: Radiation dose optimization and thoracic computed tomography. Radiol. Clin. North Am. 52(1), 1–15 (2014) 12. Barca, P., Paolicchi, F., Aringhieri, G., Palmas, F., Marfisi, D., Fantacci, M.E., et al.: A comprehensive assessment of physical image quality of five different scanners for head CT imaging as clinically used at a single hospital centre—a phantom study. PLoS One 16(1), e0245374 (2021) 13. Cody, D.D., Dillon, C.M., Fisher, T.S., Liu, X., McNitt-Gray, M.F., Patel, V.: AAPM medical physics practice guideline 1. b: CT protocol management and review practice guideline. J. Appl. Clin. Med. Phys. 22(6), 4–10 (2021) 14. State Regulatory Agency for Radiation Protection and Nuclear Safety: Regulation on the ionizing radiation protection in medical exposure. Official Gazette of Bosnia and Herzegovina (13) (2011) 15. Beganovi´c, A., Stabanˇci´c-Draguni´c, S., Odžak, S., Skopljak-Beganovi´c, A., Jaši´c, R., Sefi´cPaši´c, I.: Estimation of effective doses to patients in whole body computed tomography with automatic tube current modulation systems. In: CMBEBIH 2021: Proceedings of the International Conference on Medical and Biological Engineering, CMBEBIH 2021, April 21–24, 2021, Mostar, Bosnia and Herzegovina, p. 760-7. Springer (2021) 16. Serna, A., Ramos, D., Garcia-Angosto, E., Garcia-Sanchez, A.J., Chans, M.A., Benedicto Orovitg, J.M., et al.: Optimization of CT protocols using cause-and-effect analysis of outliers. Physica Med. 55, 1–7 (2018) 17. American Association of Physicists in Medicine: Adult Routine Chest CT Protocols Version 2.1 5/4/2016. aapmorg. (2016)
Importance of Patient Dose Evaluation and Optimization
251
18. Szucs-Farkas, Z.: Low-dose pulmonary CT angiography: reduced radiation exposure and iodine load at low tube kilovoltage. Imaging Med. 2(6), 695 (2010) 19. Rusandu, A., Ødegård, A., Engh, G., Olerud, H.M.: The use of 80 kV versus 100 kV in pulmonary CT angiography: an evaluation of the impact on radiation dose and image quality on two CT scanners. Radiography 25(1), 58–64 (2019) 20. Bhalla, A.S., Das, A., Naranje, P., Irodi, A., Raj, V., Goyal, A.: Imaging protocols for CT chest: a recommendation. Indian J. Radiol. Imaging 29(03), 236–246 (2019) 21. Martini, K., Moon, J.W., Revel, M.P., Dangeard, S., Ruan, C., Chassagnon, G.: Optimization of acquisition parameters for reduced-dose thoracic CT: a phantom study. Diagn. Interv. Imaging 101(5), 269–279 (2020) ˇ 22. Jusufbegovi´c, M., Pandži´c, A., Busuladži´c, M., Civa, L.M., Gazibegovi´c-Busuladži´c, A., Šehi´c, A., et al.: Utilisation of 3D printing in the manufacturing of an anthropomorphic paediatric head phantom for the optimisation of scanning parameters in CT. Diagnostics 13(2), 328 (2023)
A Novel X-Ray 3D Histological Method for Paraffinated Prostate Samples Santiago Laguna-Castro1,2(B) , Teemu Tolonen3 , Brian Mphande4 Jari Hyttinen1,2 , and Antti Kaipia5
,
1 Computational Biophysics and Imaging Group, Faculty of Medicine and Health Technology,
Tampere University, 33520 Tampere, Finland [email protected] 2 Faculty of Medicine and Health Technology, BioMediTech Unit, Tampere University, 33520 Tampere, Finland 3 Department of Pathology, Fimlab Laboratories, 33520 Tampere, Finland 4 Preclinical Facility, Faculty of Medicine and Health Technology, Tampere University, 33520 Tampere, Finland 5 Department of Surgery, Tampere University Hospital, 33520 Tampere, Finland
Abstract. Prostate cancer grading currently relies on the histopathological assessment of biopsies, which are analyzed as histological slides. However, the inherent two-dimensional (2D) nature of this approach, along with the limited number of slices available for clinical analysis, increases the risk of underestimating cancerous lesions. We propose that non-destructive 3D imaging methods, such as micro-Computed Tomography, can provide a more realistic and accurate means for rapid and precise characterization of prostate malignancies, thereby minimizing diagnostic discrepancies in complex cases. In this study, we focused on the technical development of an advanced 3D X-ray micro-computed tomography method (micro-CT) for imaging formalin-fixed paraffin-embedded (FFPE) rat prostate samples. This method combines a novel paraffin-clearance protocol, the use of eosin as a contrast agent, and micro-CT imaging, followed by standard hematoxylin and eosin histology. The micro-CT imaging produced a 3D model that enabled visualization and segmentation of the prostate into two well-differentiated regions: epithelium plus stroma and lumen. The CT scan was validated with hematoxylin and eosin histology. The 3D images provided volume and area measurements from the segmented parts and displayed in fine detail the extension and continuation of the glandular luminal phase. The tissue integrity was well preserved during the process, making our protocol potentially compatible with further analytical techniques. The success of this initial approach to 3D imaging the rat prostate paves the way for the application of our method to clinical samples. Keywords: Cancer · Eosin · Micro-CT · Paraffin clearance
Supported by Tampere University and Telma ja Kalevi Kronholmin Foundation. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 252–262, 2024. https://doi.org/10.1007/978-3-031-49062-0_27
A Novel X-Ray 3D Histological Method for Paraffinated Prostate
253
1 Introduction Prostate cancer is the second most common cancer type worldwide, with an incidence of 30.7% [1]. In 2020 alone, prostate cancer caused 375,304 deaths. Histopathological assessment of needle core prostate biopsies using the Gleason Grading System (GGS) has been the gold standard for staging the disease [2]. The GGS involves visually analyzing prostate glandular patterns in two-dimensional histological images and grading cancer aggressiveness from 1 to 5 based on the disorganization of glandular morphological patterns. The level of disorganization is directly correlated with cancer aggressiveness, and GGS has been a well-established histopathological method that has proven its prognostic accuracy for over 60 years. However, in classic histopathology, the number of available tissue slides is often minimized to reduce diagnosis delivery time. While this may be advantageous, it can also be a flaw when tissue patterns are tortuous, making it difficult to assess real cancer extension and type from a few slides. Additionally, the tissue is usually trimmed along one orientation, limiting sample visualization to the cutting plane. These factors increase the risk of under-staging cancer [3]. In this context, 3D imaging has been proposed as a solution to overcome the limitations of classic histopathology. Our hypothesis is that a real three-dimensional (3D) assessment of prostate glandular morphology could provide better phenotyping of cancer, minimizing the risk of misdiagnosis and aiding in the identification of features of metastatic cancers. Aim of the study. The major aim of this study is to evaluate the feasibility of our novel 3D X-ray micro-computed tomography (micro-CT) imaging method for formalin-fixed paraffin-embedded animal prostate samples before extrapolating it to prostate human samples. We also aim to assess the method’s compatibility with standard histology and implement semi-automated segmentation techniques to differentiate luminal space of the prostate from the epithelium and stroma. Previous work. The most common 3D imaging method involves reconstructing hundreds of 2D histological slices, which has proven suboptimal for clinical purposes. The mechanical and thermal stress during microtomy causes tissue expansion and deformation each time a slice is obtained. This deviation in relative orientation between the tissue surface and the cutting blade makes it challenging to obtain perfect histological parasections [4]. Consequently, 3D reconstruction based solely on slide superposition does not realistically represent samples, and further computational corrective methods are required. This approach is unfeasible for clinical practice due to its laborious nature [5].
254
S. Laguna-Castro et al.
In contrast, micro-CT is a promising alternative to histology-based 3D imaging methods because it is non-destructive and more time-efficient. Micro-CT follows the same principles as standard clinical computed tomography, allowing samples to be imaged in 3D without further processing except for the application of a contrast enhancement agent [6].
2 Materials and Methods The method developed in this study was based on the work of Busse et al. [7], which established the use of eosin as a contrast agent for whole formalin-fixed rodent kidneys. Each molecule of eosin contains a bromine atom, whose atomic number (Z = 35) specifically yields excellent contrast enhancement. Eosin primarily binds with cationic N-terminal amino groups of cytoplasmic proteins. This interaction is maximized for micro-CT by protonating tissue samples with an acid solution prior to staining. In contrast to the original protocol, our method has been designed for small formalinfixed paraffin-embedded (FFPE) samples, such as biopsies. Consequently, it incorporates a novel paraffin clearance approach as the first step. Additionally, we propose the reinclusion of the sample in paraffin before micro-CT imaging. In summary, the main steps of the proposed method are: i) paraffin clearance, ii) eosin staining, iii) micro-CT imaging, and iv) histology (see Fig. 1) for validating the micro-CT images. In order to develop the method, a section of the lateral lobe of the prostate gland from a rat (Rattus norvegicus) was obtained. The sample had been previously fixed in formalin and embedded in paraffin, and no ethical approval was applicable since it was obtained from a separate study, in accordance with the principles of the 3Rs alternatives for animal research. The use of this sample allowed for the development and optimization of the imaging and segmentation protocol, without the need for additional animal experimentation. FFPE PROSTATE SECTION
PARAFFIN CLEARANCE
EOSIN STAINING
MICRO-CT SCANNING
HISTOLOGY
Fig. 1. Developed Micro-CT Imaging Method. The protocol was designed for a formalin-fixed paraffin-embedded rat prostate section. Initial steps include the clearance of paraffin using a citrate aqueous solution at 90° in a decloaking chamber, followed by staining with eosin, which provides excellent contrast in X-ray microtomography. After the 3D imaging acquisition, standard hematoxylin and eosin histology was performed on the very same sample, with the primary purpose of validating the tomographic images.
A Novel X-Ray 3D Histological Method for Paraffinated Prostate
255
2.1 Paraffin Clearance and Eosin Staining The FFPE prostate section was detached from the histological cassette, and the excess paraffin was trimmed off with a histological blade. To clear the paraffin, the tissue was placed in a decloaking chamber for 10 min at 90 °C, soaked in 10 mM citrate pH 6.5 + 0.05% Tween-20 solution. The sample was retrieved when the temperature dropped below 50 °C; the cleared paraffin solidified on top of the citrate-aqueous solution and was manually removed. The section was then rinsed with distilled water and immediately immersed in 15 ml of 5% [vol/vol] glacial acetic acid solution in magnesium and calciumfree PBS for 48 h at 4 °C with orbital movement. For staining, the excess acidic solution was first removed with a 1-h bath in distilled water at room temperature. The eosin preparation consisted of 10 ml of a 30% [w/v] mix of eosin Y and G ([C.I. 45380], Certistain® ) in distilled water. The sample was left in the eosin for 48 h at 4 °C with continuous movement. The next step was the inclusion of the sample in paraffin according to standard methods [8], which included serial dehydration steps, xylene embedding, and leaving the sample overnight in paraffin followed by block formation.
1ml SYRINGE
EOSIN-STAINED FFPE SAMPLE
a
b
MICRO-CT MODULE ADAPTER
Fig. 2. Micro-CT Setup: a) The interior of the micro-CT scanner, where the source, sample, and detector are located, can be observed. b) To facilitate positional fixation during imaging, the sample is stained with eosin and embedded in paraffin, and then placed in a 1 mL syringe. The syringe is attached to a modular plate that can be coupled to the rotational platform of the micro-CT.
256
S. Laguna-Castro et al.
2.2 Micro-CT Setup and Imaging Embedding the sample in paraffin provides mechanical stability during micro-CT imaging, preserves the sample, and maintains its spatial disposition. When small samples are imaged in staining or preserving solutions, they can fold and turn, making it nearly impossible to preserve their exact disposition when forming the paraffin block for histology. For micro-CT acquisition (using the X-Radia microXCT-400® , ZEISS), the stained, paraffin-embedded tissue was detached from the cassette and accommodated inside a 1 mL syringe (as shown in Fig. 2). A total of 1601 projections (16-bit) were acquired during the scanning process, with electric potential and power established at 40 kV/10 W, and an exposure time of 2 s per view. Final attenuation values reached a minimum of 5000 HU, and the source/detector distances were fixed at 40/8 mm, with a magnification of 4×. 2.3 Histology After tomographic imaging, the sample was re-embedded in paraffin and processed for 2 h at 62 °C to create a new processable paraffin block. Microtomy was conducted at 5 µm slice thickness using the Leica SM2010R microtome. From each sample, a representative slice was chosen from the top, middle, and bottom regions of the tissue for standard hematoxylin and eosin (HE) staining. Finally, the histological slides were digitally scanned at 20× magnification using the NanoZoomer S60 microscope (Hamamatsu). 2.4 Analysis 3D image reorientation Prostate 3D image was manually re-oriented with the commercially available Avizo® software (2022 version, Thermo Fisher Scientific) to make the XY visualization plane of the 3D model coincident with 2D histological image, thus, the process was repeated for each 2D image. Re-orientation was divided into 3 phases: i) the tomographic XY visualization plane was rotated until framing the longest axis of the sample. ii) Visualization depth was adjusted until histology-matching structures were identified. iii) 0.01° stepped rotation in XY, XZ and YZ planes was performed until the histological image and the displayed tomosection were coincident.
A Novel X-Ray 3D Histological Method for Paraffinated Prostate
257
The prostate 3D image was manually re-oriented using the commercially available Avizo® software (version 2022, Thermo Fisher Scientific) to align the XY visualization plane of the 3D model with each 2D histological image. The re-orientation process was divided into three phases: i) The tomographic XY visualization plane was rotated until it framed the longest axis of the sample. ii) The visualization depth was adjusted until histology-matching structures were identified. iii) A 0.01° stepped rotation was performed in the XY, XZ, and YZ planes until the histological image and the displayed tomosection were coincident. Preprocessing, segmentation and measurement of the 3D image. Post-tomography image analysis was performed using Avizo® software. Initially, the image underwent median filtering. The segmentation process consisted of greyscale value thresholding for voxel selection, followed by growing and smoothing of the selection, and manual refinement. This process was repeated twice, once for each targeted segmentation material: luminal space and epithelium plus stroma. (See Fig. 3 for details.)
Fig. 3. Tomographic image segmentation workflow. The displayed steps were performed to deferentially select the epithelium plus stroma and the luminal phase of the imaged prostate section. The process was performed twice, one for each segmentation material.
258
S. Laguna-Castro et al.
3 Results and Discussion 3.1 Paraffin Clearance and Eosin Staining The sample showed no visual traces of paraffin after clearance. The intratissular paraffinclearance was inferred from the micro-CT image, which demonstrated that the eosin (an aqueous solution) infiltrated the tissue evenly, resulting in well-contrasted images. The paraffin removal in this protocol is driven by three factors. Firstly, the temperature of the clearing solution, set at 90 °C, which is beyond the melting point of paraffin (60 °C). Secondly, the density gradient established between the paraffin and the citrate-aqueous media, which forces the paraffin outwards from the tissue and into the deparaffination media. Lastly, the presence of Tween® in the solution acts as a surfactant and facilitates the escape of paraffin out of the cell membranes. 3.2 Micro-CT Imaging, Segmentation and Measurements The final image quality was good, as evidenced by the final scan resolution of 5.6419 µm/pixel and a signal-to-noise ratio measurement of 39.38 (see Fig. 4 for details).
Fig. 4. Prostate vs background grey scale values histogram. Data was acquired from of two representative region-of-interest ROIs (330 × 390 × 1330 µm). Notice two well differed local maxima in the tissue histogram (white). These peaks were empirically associated with with the luminal phase (left peak), and the rest of the tissue (right peak). Signal-to-noise ratio (SNR), was calculated as the mean value of greyscale value of the tissue ROI image divided by the standard deviation (std).
The implemented segmentation pipeline resulted in consistent differentiation of the prostate tissue from the luminal space (see Fig. 5). However, while the epithelium was visually distinguishable on the micro-CT tomosections, segmentation of the epithelium based on basic grey values thresholding was insufficient to fully isolate it from the stromal tissue and highlights the analytical limitations when traditional segmentation technique are applied. From our perspective, this denotes the need to incorporate high-performance computational methods like deep neural networks training and classification during for
A Novel X-Ray 3D Histological Method for Paraffinated Prostate
259
segmentation phase. Geometrical measurements of the final 3D model (see Table 1) revealed a higher surface density or area/volume ratio for the luminal phase, which appeared as network of cavities in the 3D image. On the other hand, the epithelium and stroma presented higher volume and area values when compared with the lumen. This result was expected to some extent since the only the epithelium tends to occupy a larger volume than the lumen, as it is composed of multiple layers of different cell types. Parallelly, the surface density or volume/area ratio in prostatic tissue is dependent on several factors such as age, hormonal status, and secretory activity. In pathological conditions like prostate cancer, there may be an imbalance between the epithelium and lumen, leading to architectural disruption of glandular structures and to an irregular distribution of epithelial and luminal volumes, as described in the introduction chapter.
Fig. 5. 3D model of segmented rat prostate section. In a, epithelium (yellow, left) and lumen(blue, right) can be observed from three different planes. In b, perspective view of both tissue types are displayed. It is to be pointed out the disposition of the luminal phase, which is presented as an almost continuous and sinuous cavity.
260
S. Laguna-Castro et al.
Table 1. Absolute values of volume, area and surface density of the computationally segmented rat prostate sample. Material
Volume [µm3 ]
Area [µm2 ]
Surface density*
Epithelium + Stroma
4.83 × 109
2.27 × 108
21.30
Lumen
2.78 × 108
8.28 × 107
33.60
* Surface Density = volume/area
Fig. 6. Comparison between micro-CT tomosections and subsequent HE histology. Micro-CT tomosections are representative of the top, middle and bottom of the sample along the z-axis (a–c). Histology slices (d–l) are corresponding with the tomosections and displayed at different magnifications. The micro-ct images were manually re-oriented to be coincident with the visualization plane of the histological slides.
3.3 Micro-CT vs Histology The final validation of the micro-CT images with the corresponding histological slides resulted in a satisfactory level of concordance in terms of morphological features. Notably, the quality of the histology was found to be fairly optimal after the microCT scanning procedure, with no significant artifacts related to tissue damage or tearing. However, it is worth noting that the luminal space in the histological slides appeared empty (see Fig. 6).
A Novel X-Ray 3D Histological Method for Paraffinated Prostate
261
4 Conclusions In summary, our developed method has been demonstrated to be fully compatible with the use of formalin-fixed paraffin-embedded (FFPE) samples and standard histology, while yielding robust quality in both the tomographic and histological images. This approach represents a significant improvement in 3D histology compared to more traditional methods that rely on the alignment of 2D histological slides, which often fail to provide consistent and realistic representation of tissue in three dimensions. However, it is important to note that the main limitation of this study lies in the selected segmentation approach. As the segmentation was based on basic thresholding and manual adjustments, it was unable to differentiate between stroma and epithelium. This finding highlights the need for more sophisticated computational tools and deep neural networks to maximize the segmentation possibilities of micro-CT images. The primary focus of our forthcoming research will be to address the current limitation by developing advanced segmentation strategies that can fully capture the complexity of tissue architecture. Special emphasis will be placed on devising sophisticated approaches capable of accurately differentiating between stroma and epithelium, utilizing cutting-edge computational tools and deep neural networks to improve the segmentation performance of micro-CT images. Acknowledgements. We would like to express our sincere gratitude to the lab technicians who contributed to the success of this study. Their tireless efforts and expertise were invaluable to the completion of this project. Furthermore, we would like to acknowledge the generous sponsorship provided by the Telma ja Kalevi Kronholmin Foundation. Their financial support allowed us to carry out the necessary experiments and data analysis. We are also grateful to our home institution, the Faculty of Medicine and Health Technology, who provide us with all the infrastructure to complete this project. Finally, we would like to thank our colleagues for their helpful discussions and feedback throughout the course of this study.
References 1. World Health Organization. Cancer Today. www.Iarc.fr, www.gco.iarc.fr/today/home. Last accessed 21 Feb 2023 2. Egevad, L., et al.: International Society of Urological Pathology (ISUP) grading of prostate cancer. Am. J. Surg. Pathol. 40(6), 858–861 (2016). https://doi.org/10.1097/PAS.000000000 0000642 3. Barner, L.A., Glaser, A.K., Mao, C., et al.: Multiresolution nondestructive 3D pathology of whole lymph nodes for breast cancer staging. J. Biomed. Opt. 27(3), 036501 (2022). https:// doi.org/10.1117/1.JBO.27.3.036501 4. Hillman, H.: Limitations of clinical and biological histology. Med. Hypotheses 54(4), 553–564 (1999). https://doi.org/10.1054/mehy.1999.0894 5. Song, Y., Treanor, D., Bulpitt, A.J., Magee, D.R.: 3D reconstruction of multiple stained histology images. J. Pathol. Inform. 4(7). https://doi.org/10.4103/21533539.109864 6. Laguna-Castro, S.: Establishment of new micro-computed tomography methodologies. [MSc Thesis]. http://urn.fi/URN:NBN:fi:tuni-202010287652
262
S. Laguna-Castro et al.
7. Busse, M., Müller, M., Kimm, M.A., et al.: Three-dimensional virtual histology enabled through cytoplasm-specific X-ray stain for microscopic and nanoscopic computed tomography. PNAS 115(10), 2293–2298 (2018). https://doi.org/10.1073/pnas.1720862115 8. Kim Suvarna, S., Layton, C., Bancroft, J.D.: Bancroft’s Theory and Practice of Histological Techniques, 8th edn, pp.78–83. Oxford, Elsevier, (2019). ISBN 978-0-7020-6864-5
Review of MRI Reporter Genes in Oncology 2 , Zerina Kali´ ˇ Adna Softi´c1(B) , Ivana Ceko c2 , Nejla Piri´c2 , Emina Mrdanovi´ c1 , 1 and Elma Imamovi´c 1 Verlab Institute, 71000 Sarajevo, Bosnia and Herzegovina
[email protected] 2 International Burch University, Ilidža, 71210 Sarajevo, Bosnia and Herzegovina
Abstract. Molecular imaging aims to detect molecular events in entire organisms. Even though it is not possible to directly detect the genes and proteins located in intracellular compartments, it is possible to detect them via reporter genes that encode for the specific proteins. The reporter gene has gained importance in diagnostics, enabling monitoring of gene expression patterns and in vivo monitoring of cell survival, proliferation, migration and differentiation. As cancer is the one of the most significant burdens for both population health and healthcare systems, it is of high importance to devote significant efforts towards timely diagnosis. The MRI reporter genes can be used for this purpose. For the purposes of this work, the following databases were searched: ScienceDirect, PubMed and Google Scholar. The inclusion criteria is the publications containing keywords such as MRI, MRI reporter gene, MRI reporter gene cancer, Molecular-Genetic Imaging, and reporter gene imaging, language of publication is English, articles including in vivo monitoring released between 2000 and 2023, and full-text journal publications, conference publications, standards, guidelines, and books. Applying the previous criterion 24 articles were found and analyzed. This review will present some of the MRI reporter genes that are used in oncology with an emphasis on most recent trends. Keywords: MRI reporter genes · Oncology · Cancer
1 Introduction Magnetic Resonance Imaging (MRI), a non-invasive imaging technology, produces three dimensional detailed anatomical images. It is frequently used for disease detection, diagnosis, and treatment monitoring [1]. The sensitivity and specificity of MRI was increased by the addition of reporter genes. When introduced into target cells - brain tissues, cancer and circulating white cells, reporter genes produce a protein receptor or enzyme that binds, transports or traps a subsequently injected imaging probe. Reporter genes imaging is an effective tool for studying molecular activities in living organisms. MRI reporter genes are genes that encode for proteins that can be imaged using magnetic resonance imaging (MRI) techniques. These genes have the potential to be powerful tools for non-invasive monitoring of gene expression and cell tracking in living organisms. In oncology, MRI reporter genes are being developed to improve the detection and treatment of cancer [2]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 263–269, 2024. https://doi.org/10.1007/978-3-031-49062-0_28
264
A. Softi´c et al.
The first MRI reporter gene utilized in the past was creatine kinase (CK), an enzyme that catalyzes the conversion of adenosine triphosphate (ATP) to adenosine diphosphate (ADP) while simultaneously producing phosphocreatine (PCr). This process can be detected by 31 P Magnetic Resonance Spectroscopy (MRS) [3]. Reporter genes for MRI may be used to precisely track the cell delivery in cell therapy, analyze the therapy effect of gene delivery, and monitor tissue/cell-specific microenvironments. By visualizing the levels of exogenous or endogenous gene expression, specific signal transduction pathways, nuclear receptor activities, or protein–protein interactions, MRI reporter gene imaging can longitudinally monitor the processes (e.g., cell delivery, gene expression, et al.) in living organisms [4, 5]. In order to measure expressed reporter gene protein, various detection methods are used. These methods include luminescence, absorbance and fluorescence [6]. Commonly used reporter genes for MRI usually include genes encoding the enzyme (e.g., tyrosinase and β-galactosidase), the receptor on the cells (e.g., transferrin receptor), and endogenous reporter genes (e.g., ferritin reporter gene) [7–9]. MRI can visualize functional physiologic changes or biochemical changes, as well as anatomic morphologic changes using radiolabeled metabolic substances, such as glucose, amino acids, or nucleotides. Classified MRI approaches on the basis of interactions at the molecular level: the expression of surface receptors that allow the binding of specific MRI contrast agents; the enzyme-based cleavage of functional groups that block water (proton) exchange, protein binding, or MRI contrast agents; and the expression of para- and antiferromagnetic proteins involved with iron metabolism, such as tyrosinase and ferritin [10]. Despite the potential of MRI reporter genes, there are still obstacles to their broad usage. One significant problem is getting high levels of reporter gene expression in target cells, which is required for precise imaging. Moreover, the sensitivity of MRI methods may restrict the capacity to identify modest levels of gene expression. Finally, MRI reporter genes have the potential to be useful tools for tracking cancer progression and therapy. Aim of this paper is to present all MRI reporter genes which are used in oncology with the emphasis on most recent trends. Further study is needed, however, to improve their expression and imaging capabilities, as well as to establish their therapeutic value. Therefore, the purpose of our review is to present the current state of research on MRI reporter genes in oncology, so it can help to advance the field and identify areas for further study.
2 Methodology ScienceDirect, PubMed and Google Scholar were reviewed to find as much information as possible about reporter genes in oncology analyses. The following search criteria were used to limit the research: 1) publications containing keywords: Magnetic Resonance Imaging, MRI reporter gene, MRI reporter gene cancer, Molecular-Genetic Imaging, and Reporter gene imaging 2) language of publication: English 3) publication time: 2000–2023 4) full-text journal publications, conference publications, standards, guidelines, books.
Review of MRI Reporter Genes in Oncology
265
A summary of the most commonly studied genes that have been used as MRI reporters in oncology is presented in this paper. There are two categories of reporter genes: those that can be visualized without the need to administer a substrate in the form of a contrast agent and those that rely on substrate injection to provide contrast.
3 Results and Discussion By conducting a thorough examination of scientific literature and scrutinizing various methodologies for creating Magnetic Resonance Imaging (MRI) reporters, three distinct contrast mechanisms/agents were identified, namely T2 or T2*, T1, Chemical Exchange Saturation Transfer (CEST). This review comprehensively presents the reporters and sensors utilizing the above-mentioned contrast mechanisms and elaborates on their individual features. After the screening process, 10 reporter genes were found to be most suitable for the intended purpose. In a study by Qin et al. MCF-7-TYR human breast cancer cells have been transfected with a plasmid encoding tyrosine (TYR) and untransfected MCF-7 cells served as the study’s negative controls. MCF-7-TYR tumors achieved much larger signals and tumorto-background contrasts than MCF-7 tumors in in vivo MRI imaging studies, among others such as PET and PAI [11]. The finding that increased levels of tyrosinase activity in melanotic melanomas led to stronger signal intensity in T1-weighted MR images of patients led Vandsburger et al. to investigate whether overexpressing tyrosinase could be used as a way to track cell survival through MRI [12]. Initial studies showed that introducing human tyrosinase to murine fibroblasts and human kidney cells resulted in higher MRI signal intensity. Two subsequent studies by Alfke et al. and Paproski et al. confirmed these findings by inducing the expression of tyrosinase in human breast cancer cells [13, 14]. Arena et al. used the Lac Z gene as an MRI reporter to track mouse melanoma cell proliferation by administering a gadolinium-based contrast agent that is cleaved by the β-galactosidase enzyme expressed by Lac Z. Cancer cells expressing Lac Z show higher contrast on T1 weighted MRI compared to control [15]. Additionally, in the study by Cui et al., the interaction between β-gal and a staining salt generates strong hypointensity on T2* weighted images in LacZ-expressing tumor cells in the presence of ferric ions [16]. According the studies by Zhou et al. from 2020 and Yang et al. from 2016, that we analyzed, ferritin heavy chain (Fth) is a promising magnetic resonance (MR) reporter gene for imaging tumors due to its ability to reduce signal intensity without fading over time and minimal cellular toxicity [17, 18]. Fth overexpression induces iron deficiency mechanisms, such as the expression of transferrin receptor (TfR), which makes it a good MR endogenous contrast agent. Moreover, Fth overexpression can enhance sitespecific drug delivery by inducing the expression of TfR, which is overexpressed on cancer cells compared to normal cells [19]. Liposomes are a widely studied carrier that can be modified with transferrin (Tf) to achieve targeted drug delivery to cancer cells overexpressing TfR. Encapsulation of the broad-spectrum antitumor drug within Tfmodified liposomes (Tf-LPD) can improve therapeutic efficacy while reducing adverse effects [20]. A study by Sun et al. from 2021 aimed to develop an MRI monitoring system for early detection of potential malignant transformation of stem cells. The system required
266
A. Softi´c et al.
incorporation of an MRI reporter gene and its expression in a tumor-specific manner. Fth1 was identified as an ideal MRI reporter gene, but its continuous expression made it unsuitable for detecting malignant transformation. The study modified Fth1 by adding the tumor-specific promoter PEG3 upstream, and successfully transferred it into stem cells using a recombinant lentivirus. The modified gene remained silent in normal stem cells but turned on when stem cells transformed into tumor cells, enabling MRI detection. This system could detect unexpected tumorigenicity in stem cell-based therapies and monitor the elimination of malignantly transformed cells in combination with a therapeutic gene [21]. Several studies conducted on cancerous rodent models employed Fth overexpression as a means to visualize the proliferation of murine melanoma cells upon implantation in close proximity to significant lymph nodes [22]. Additionally, Fth overexpression was utilized to monitor the proliferation of human breast cancer and rat glioma cells following subdermal implantation in mice and rats [23]. Bartelle et al. developed a system in which a cluster of biotins is displayed on cell membranes by a synthetic “Biotag” reporter. The transgene is entirely self-contained and can be employed as a vascular reporter system with any endothelial promoter, in contrast to earlier biotinylation methods. Transgenic Ts-Biotag mice were generated with a minimal reporter for Tie2. The potential of this technique for vascular imaging of genetic processes in mice was shown by the in vivo imaging of Tie2 expression in TsBiotag mice, in both embryonic and adult models of vascular development [24]. Another study by Xavierselvan et al. demonstrated that imaging of tumor vasculature can be a valuable tool for preclinical cancer research. Ergo, the Biotag reporter could be used and tested for vascular MRI imaging of tumors [25]. The bacterial gene MagA can make mammalian cells magnetic and suitable for tracking through magnetic resonance imaging (MRI). In a mouse model, MagA-expressing tumors showed increased and higher quality MR contrast compared to tumors expressing a modified ferritin system lacking iron response element regulation. The results suggest that MagA expression can be used for monitoring cell growth and differentiation with the potential for in vivo detection of reporter gene expression using MRI [26]. The divalent metal ion transporter (DMT1) and a biotinylated cell surface protein (BAP-TM) have shown the best contrast in vivo among positive contrast agents for T1weighted MRI. DMT1 led to a 75% increase in image contrast in glioma tumors, but with significant enhancement in surrounding normal tissue and variation in persistence of enhancement between tumor types [27]. b-Galactosidase is particularly useful as an MRI reporter gene because of its low background level. The presence of galactosidase will cleave the galactose and release the Gd31 contrast agent [28]. Furthermore, the nonmetallic, biodegradable MRI reporter gene encoding lysine-rich protein (LPR) reduces MRI signal intensity through the rapid transfer of amide proton in LRP to a water proton, which produces a saturation transfer contrast in solution. The different signal intensity change for an LRP-expressing xenograft in comparison to a control xenograft in mice confirmed their potential as a reporter [4]. A study by Farrar et al. showed that incorporation of the LRP reporter gene into G47, an oncolytic virus originating from herpes simplex, had no effect on the viral replication or therapeutic effectiveness. These findings are significant because they
Review of MRI Reporter Genes in Oncology
267
suggest that the LRP gene can potentially be used as a reporter for the real-time detection of viral spread [29]. Tyrosinase and ferritin are metalloprotein-based MRI reporter genes, the former playing a role in melanin biosynthesis, resulting in changes in MRI signals in melanocytes and melanoma cells [13, 30]. The latter’s iron-sequestering properties make it a promising MRI reporter. Cells transfected with a ferritin reporter gene have the ability to overexpress it, causing the capture of extracellular/endogenous iron, and formation of superparamagnetic crystalline iron, which leads to an observable MRI contrast [31]. Nystrom et al. have developed an imaging system using an organic anion-transporting polypeptide 1b3 (oatp1b3) as an MRI reporter gene that enables sensitive threedimensional detection of viable cancer cells in live mice, providing high-resolution imaging and longitudinal tracking of metastatic spread to multiple lymph nodes and different organ systems in individual animals. This system could help improve our understanding of metastasis and aid in the development of new cancer therapies [32, 33]. MRI reporter genes and their details are shown in Table 1. Table 1. Short review on MRI reporter genes in oncology Reporter gene
Contrast mechanism
Substrate
Cancer type
References
Tyrosinase
T2/T1
Endogenous iron
Human breast cancer
[11, 12]
Ferritin
T2
No substrate
Liver cancer; Malignant stem cells
[17, 18, 21]
Transferrin receptor
T2
Tf-MIONs
Solid tumor
[19]
Lac Z
T1/T2*
S-Gal
Human breast cancer; Mouse melanoma
[15, 16]
β-galactosidase
T1
Gadolinium ion
Human breast cancer; Mouse melanoma
[15, 16, 28]
Biotag
T1
Gadolinium based substrate
Tumor vasculature
[24, 25]
MagA
T2
Iron supplement
Melanoma
[26]
DMT1
T1
Manganese-based substrate
Melanoma
[27]
LRP
CEST
N/A
Rat glioma
[29]
Oatp1b3
T1
Gadoxetate disodium Prostate cancer
[32, 33]
268
A. Softi´c et al.
References 1. National Institute of Biomedical imaging and bioengineering: Magnetic Resonance Imaging (MRI). National Institute of Biomedical Imaging and Bioengineering (2018). www.nibib.nih. gov/science-education/science-topics/magnetic-resonance-imaging-mri 2. Budinger, T.F., Jones, T.: 1.01—History of Nuclear Medicine and Molecular Imaging. In: Brahme, A. (ed.) ScienceDirect, pp. 1–37. Elsevier (2014) www.sciencedirect.com/science/ article/pii/B9780444536327001015 3. Gilad, A.A., et al.: MRI reporter genes. J. Nucl. Med. Off. Publ. Soc. Nucl. Med. 49(12), 1905–1908 (2008). https://doi.org/10.2967/jnumed.108.053520 4. Kang, J.H., Chung, J.K.: Molecular-genetic imaging based on reporter gene expression. J. Nucl. Med. 49(Suppl. 2), 164S-179S (2008). https://doi.org/10.2967/jnumed.107.045955 5. Harney, A.S., Meade, T.J.: Molecular imaging of in vivo gene expression. Future Med. Chem. 2, 503–519 (2010). https://doi.org/10.4155/fmc.09.168 6. Yang, C., Tian, R., Liu, T., Liu, G.: MRI reporter genes for noninvasive molecular imaging. Molecules 21(5), 580 (2016). https://doi.org/10.3390/molecules21050580. PMID: 27213309; PMCID: PMC6273230 7. He, X., Cai, J., Liu, B., Zhong, Y., Qin, Y.: Cellular magnetic resonance imaging contrast generated by the ferritin heavy chain genetic reporter under the control of a Tet-On switch. Stem Cell Res. Ther. 6 (2015). https://doi.org/10.1186/s13287-015-0205-z 8. Cai, Y., et al.: Enhanced magnetic resonance imaging and staining of cancer cells using ferrimagnetic H-ferritin nanoparticles with increasing core size. Int. J. Nanomed. 10, 2619– 2634 (2015) 9. Lin, X., et al.: Chimeric ferritin nanocages for multiple function loading and multimodal imaging. Nano Lett. 11, 814–819 (2011). https://doi.org/10.1021/nl104141g 10. Gilad, A.A., Winnard, P.T., Jr., Van Zijl, P.C., Bulte, J.W.: Developing MR reporter genes: promises and pitfalls. NMR Biomed. 20, 275–290 (2007) 11. Qin, C., Cheng, K., Chen, K., et al.: Tyrosinase as a multifunctional reporter gene for Photoacoustic/MRI/PET triple modality molecular imaging. Sci. Rep. 3, 1490 (2013). https://doi. org/10.1038/srep01490 12. Vandsburger, M.H., Radoul, M., Cohen, B., Neeman, M.: MRI reporter genes: applications for imaging of cell survival, proliferation, migration and differentiation. NMR Biomed. 26(7), 872–884 (2012). https://doi.org/10.1002/nbm.2869 13. Alfke, H., et al.: In vitro MR imaging of regulated gene expression. Radiology 228(2), 488–492 (2003) 14. Paproski, R.J., Forbrich, A.E., Wachowicz, K., Hitt, M.M., Zemp, R.J.: Tyrosinase as a dual reporter gene for both photoacoustic and magnetic resonance imaging. Biomed. Opt. Express 2(4), 771–780 (2011) 15. Arena, F., et al.: β-Gal gene expression MRI reporter in melanoma tumor cells. Design, synthesis, and in vitro and in vivo testing of a Gd(III) containing probe forming a high relaxivity, melanin-like structure upon β-Gal enzymatic activation. Bioconjug. Chem. 22(12), 2625–635 (2011). https://doi.org/10.1021/bc200486j 16. Cui, W., Liu, L., Kodibagkar, V.D., Mason, R.P.: S-Gal, a novel 1H MRI reporter for betagalactosidase. Magn. Reson. Med. 64(1), 65–71 (2010). https://doi.org/10.1002/mrm.22400 17. Zhou, J., et al.: Dual-effect of magnetic resonance imaging reporter gene in diagnosis and treatment of hepatocellular carcinoma. Int. J. Nanomed. 15, 7235–7249 (2020). https://doi. org/10.2147/IJN.S257628 18. Yang, Y., Gong, M.F., Yang, H., et al.: MR molecular imaging of tumours using ferritin heavy chain reporter gene expression mediated by the hTERT promoter. Eur. Radiol. 26(11), 4089–4097 (2016). https://doi.org/10.1007/s00330-016-4259-9
Review of MRI Reporter Genes in Oncology
269
19. Feng, Y., et al.: Efficiency of ferritin as an MRI reporter gene in NPC cells is enhanced by iron supplementation. J. Biomed. Biotechnol. 2012, 434878 (2012). https://doi.org/10.1155/ 2012/434878 20. Moghimipour, E., Rezaei, M., Kouchak, M., et al.: A mechanistic study of the effect of transferrin conjugation on cytotoxicity of targeted liposomes. J. Microencapsul. 35(6), 548– 558 (2018). https://doi.org/10.1080/02652048.2018.1547325 21. Sun, J., Huang, J., Bao, G., et al.: MRI detection of the malignant transformation of stem cells through reporter gene expression driven by a tumor-specific promoter. Stem Cell Res. Ther. 12, 284 (2021). https://doi.org/10.1186/s13287-021-02359-w 22. Choi, S.H., et al.: Imaging and quantification of metastatic melanoma cells in lymph nodes with a ferritin MR reporter in living mice. NMR Biomed. 25(5), 737–745 (2012) 23. Kim, H., Cho, H., Choi, S., Woo, J., Moon, W.: In vivo imaging of tumor transduced with bimodal lentiviral vector encoding human ferritin and green fluorescent protein on a 1.5T clinical magnetic resonance scanner. Cancer Res. 70(18), 7315–7324 (2010) 24. Bartelle, B.B., et al.: Novel genetic approach for in vivo vascular imaging in mice. Circul. Res. 110(7), 938–947 (2012). https://doi.org/10.1161/CIRCRESAHA.111.254375 25. Xavierselvan, M., Singh, M.K.A., Mallidi, S.: In Vivo tumor vascular imaging with light emitting diode-based photoacoustic imaging system. Sensors 20, 4503 (2020). https://doi. org/10.3390/s20164503 26. Rohani, R., Figueredo, R., Bureau, Y., et al.: Imaging tumor growth non-invasively using expression of MagA or modified ferritin subunits to augment intracellular contrast for repetitive MRI. Mol. Imaging Biol. 16, 63–73 (2014). https://doi.org/10.1007/s11307-0130661-8 27. Bartelle, B.B., Szulc, K.U., Suero-Abreu, G.A., Rodriguez, J.J., Turnbull, D.H.: Divalent metal transporter, DMT1: a novel MRI reporter protein. Magn. Reson. Med. 70(3), 842–850 (2013). https://doi.org/10.1002/mrm.24509 28. Louie, A.Y., et al.: In vivo visualization of gene expression using magnetic resonance imaging. Nat. Biotechnol. 18(3), 321–325 (2000). https://doi.org/10.1038/73780. PMID: 10700150 29. Farrar, C.T., et al.: Establishing the lysine-rich protein CEST reporter gene as a CEST MR imaging detector for oncolytic virotherapy. Radiology 275(3), 746–754 (2015). https://doi. org/10.1148/radiol.14140251 30. Yang, C., et al.: MRI reporter genes for noninvasive molecular imaging. Molecules (Basel, Switzerland) 21(5), 580 (2016). https://doi.org/10.3390/molecules21050580 31. Brindle, K.M.: Gene reporters for magnetic resonance imaging. Trends Genet. 38(10), 996– 998 (2022). https://doi.org/10.1016/j.tig.2022.05.006 32. Nyström, N.N., McRae, S.W., Martinez, F.M., Kelly, J.J., Scholl, T.J., Ronald, J.A.: A genetically encoded magnetic resonance imaging reporter enables sensitive detection and tracking of spontaneous metastases in deep tissues. Can. Res. 83(5), 673–685 (2023). https://doi.org/ 10.1158/0008-5472.CAN-22-2770 33. Lochrin, S.E., Turkbey, B., Gasmi, B., et al.: Pilot study of gadoxetate disodium-enhanced mri for localized and metastatic prostate cancers. Sci. Rep. 11, 5662 (2021). https://doi.org/ 10.1038/s41598-021-84960-w
Image Registration Techniques for Independent Acquisitions of Cone Beam Computed Tomography Volumes Diletta Pennati1,3,4(B) , Leonardo Manetti3,4 , Ernesto Iadanza2 , and Leonardo Bocchi1,4 1 Department of Information Engineering, University of Florence, Florence, Italy
[email protected]
2 Department of Medical Biotechnologies, University of Siena, Siena, Italy 3 Epica Imaginalis, Via Rodolfo Morandi 15, Sesto Fiorentino, Italy 4 EIDO Lab, Florence, Italy
Abstract. This work is centered on the alignment of a volumetric dataset acquired through Cone Beam Computed Tomography (CBCT) technology. Monomodal registration is useful for comparing different acquisitions of the same anatomical district for monitoring a pathology progression or regression, as well as for stitching together CBCT consecutive volume segments, usually when a large region of interest does not in fit the device field of view. Several methods were studied, both intensity and feature-based. Gradient-free techniques and evolutionary algorithm class, in particular genetic algorithms, were investigated. Results were analyzed to establish which approach is more efficient and accurate. Convergence speed represents a known issue of this evolutionary algorithms that was handled through the choice of an adequate stop criterion. Results were presented over a dataset, where a known rigid transformation matrix is applied, with the aim of comparing the estimated transformations with the actual ones.
1 Introduction In the medical field, the possibility of having an instrument for visualizing different information on a unique volumetric dataset has increasingly become a powerful tool for clinicians to provide a more comprehensive patient diagnosis. For this purpose, images must be accurately co-registered to align anatomical structures in both datasets as to perform a proper comparison. Image registration has several applications and can involve more than one imaging technique (multimodal registration). Multimodal image fusion can be used to compensate for the lack of morphological or functional information in one of the individual imaging modalities. Instead, when a single modality is involved, the purpose of image registration is to observe the same region of interest under different conditions and/or at different time points. Image fusion has demonstrated its advantages in Computer-Assisted Orthopaedic Surgery (CAOS), a real-time navigation guidance system that assists surgeons intraoperatively [8]. Medical image registration methods can be grouped based on some criteria (Fig. 1): the number of different modalities involved; the subjects involved: a co-registration can © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 270–277, 2024. https://doi.org/10.1007/978-3-031-49062-0_29
Image Registration Techniques for Independent Acquisitions
271
be performed on the same patient at different times as well as between different patients; alignment with a reference image from a medical atlas can also be useful for comparisons with pre-classified images; image dimensionality: co-registration can be performed both at single slice level and at three dimensional level; the type of registration basis: extrinsic methods involve extraneous objects introduced into the image space, intrinsic methods are based on the image information generated by the patient himself; the nature of transformation applied to the moving image; the transformation domain: depending on the application, registration at the global level might be unnecessary, whereas focusing on a limited region of interest might be sufficient; the degree of interaction between the user and the registration system. These and other criteria are not analyzed in detail in this paper. In recent years, several reviews and surveys have been published that provide a detailed framework for the co-registration of medical images [4, 5, 7, 9].
Fig. 1. Classification criteria of medical image registration
This work focuses on the subcategory of the so-called monomodal image registration, specifically using images obtained with Cone Beam Computed Tomography (CBCT) technology. This category of image registration can be performed for a variety of purposes. Because the field of view of the CBCT detector is generally limited to a small region of interest, two or more consecutive acquisitions need to be performed when a large volume has to be analysed. To obtain a unique volume, overlapping slices must be registered, in order to have the correct matching between identical features. Subvolumes stitching can also be useful when continuous patient bed movement is not possible and the patient must be manually positioned prior to the second acquisition. Monomodal tasks are also well suited for intervention verification, monitoring the progression or regression of pathology, and many other applications [9]. In these cases, the assumption of coplanarity of the two partial volume slices must be dropped to obtain consistent results, since the acquisitions are performed at different times and likely with different
272
D. Pennati et al.
positioning. The present work will be focused on the first listed application, i.e. registration of consecutive CBCT scans covering large regions of interest. However, it is possible to extend the present algorithm to other applications where slice coplanarity cannot be assumed by increasing the number of degrees of freedom, including affine or non-rigid transformations. In the next sections, the image registration process is described in its main steps and the most commonly used techniques are listed. In the results section, an example of application to CBCT images using one of the techniques described in the previous section is presented.
2 Materials and Methods Image registration can actually be viewed as an optimization process, as shown in Fig. 2. The cost function to be minimized describes the degree of similarity between the image used as reference and the moving image, and depends on the transformation matrix applied to the moving image, which has been properly interpolated [5]. The image registration process follows two main approaches: intensity-based methods and feature-based methods. In intensity-based methods, pixels/voxels grey level intensity is used to compare the aligned images. The simplest similarity metric is the sum of squared differences of pixel intensity values. However, this measure is highly affected by noise; moreover, it is not possible to assume identical intensity levels between two consecutive scans, even if the X-ray acquisition parameters are the same. To overcome this limitation, Mutual information (MI) is considered the gold standard for the intensity-based approach because of its robustness, accuracy and universality. MI is a statistical measure that calculates the mutual dependence of the underlying image intensity distributions by exploiting the nonlinear correlations between them. MI assumes that the co-occurrence of the most probable values in the two datasets is maximised when they are matched. The problem with similarity criteria based on information theory is that they do not have a direct mapping to spatial information, and therefore there is always the possibility that important anatomical information will be lost [5]. For this reason, hybrid methods that combine spatial information with mutual information have been proposed to improve the accuracy of image registration algorithms [1]. Feature-based methods use artificial external or anatomical internal markers for image alignment. Since the CBCT technology produces good images in terms of spatial resolution, an intrinsic feature approach can be considered without using external markers. Several feature detection methods have been developed and compared [10, 11]. All methods studied are based on a multi-resolution approach: each level, called octave, is the downsampled version of the previous one. In SIFT (Scale-invariant Feature Transform) and SURF (Speeded Up Robust Features) algorithms, images belonging to the same octave are smoothed with a Gaussian kernel with increasing sigma: however, this results in smoothing of both noise and high frequency elements, which can lead to a loss of detail. For this reason, the Accelerated-KAZE (AKAZE) method developed by Alcantarilla et al. [2] is adopted in this work. This algorithm is based on Fast Explicit Diffusion (FED), a nonlinear diffusion filtering technique that removes noise without losing essential details. The AKAZE detector is based on the determinant of the Hessian matrix. Maxima of detector responses at spatial locations are picked up as feature
Image Registration Techniques for Independent Acquisitions
273
Fig. 2. Image registration process
points. A Modified-Local Difference Binary (M-LDB) descriptor was used that exploits gradient information from nonlinear scale space. AKAZE features are invariant to scale, rotation, and limited affine transformations. Once these keypoints are extracted and correctly matched, the geometric distance between the corresponding keypoints can be used as a similarity metric to be minimized. Concerning the optimization algorithm, gradient-based methods are also commonly used by precompiled registration libraries [5]. A variant of gradient descent, such as Regular Step Gradient Descent, is used to prevent too large steps. At each iteration, this optimizer takes a step along the direction of the metric derivative. If the direction of the derivative changes abruptly, the optimizer assumes that a local extremum has been passed and reacts by reducing the step length by a relaxation factor. In this work, a population-based approach, specifically the Particle Swarm Optimization (PSO) algorithm, has been used. In this method, inspired by the swarming behaviour of biological populations (Fig. 3), the particles (the solutions) traverse the input space of the cost function (the chosen similarity metric) to find the global minimum. The particles’ movements are guided by their own best known position in the search space as well as the best known position of the entire swarm [6]. PSO is an iterative procedure in which a particle moves in each iteration with a velocity vector that is a function of the best position (position with the lowest objective function value) found by that particle and the best position found by all particles so far. The particle position update is expressed by the following Equation: vik (t + 1) = wvik (t) + c1 α1 (pbestik − xik (t)) + c2 α2 (gbest k − xik (t)) xik+1 (t + 1) = xik (t) + vik+1 (t + 1)
(1)
where vik is the velocity of the ith particle at the kth iteration, and xik is the current solution (or position) of the ith particle at the kth iteration. c1 , c2 are positive constants, and α 1 , α 2 are two random variables with a uniform distribution between 0 and 1. In this equation, w is the inertia weight which shows the effect of the previous velocity vector on the new vector.
274
D. Pennati et al.
Fig. 3. Particle position update according to swarm algorithm
The advantages of PSO are insensitivity to scaling of design variables, the derivative freedom and ease of parallelization for concurrent processing. The disadvantages of PSO are that it is easy to fall into a local optimum in high-dimensional space and the low convergence rate of the iterative process [3]. In the next section, the PSO algorithm has been applied to find the optimal transformation matrix between a set of CBCT reference images and a set of CBCT images with a known transformation applied. All discrepancies between the estimated parameters and the expected misalignment are analyzed and commented.
3 Results and Discussion A previously archived dataset of a veterinary patient was used for this work. This dataset consists of two consecutive subvolumes that have to be stitched, acquired with the following parameters: 80 kV, 45 mA, 6 ms X-ray tube values, 0.25 mm slice thickness, image size = 701 × 701 pixels. A section of 10 slices where the two subvolumes overlap was extracted. A known transformation matrix was applied to one of the two sets of 10 slices, in order to compare the estimated result with the expected ones. A 2D-to-2D registration was performed between each pair of slices. Since the patient under study was not moved between the two consecutive scans, coplanarity between the slices was assumed, so only a rigid 2D transformation was allowed, i.e., translation along the x- and y-axes and rotation around the z-axis. For the 10 image pairs, 8 different experiments were performed, resulting in a total of 80 experiments (Table 1). This combination of values represents the expected result of the optimization process. Keypoints extraction was performed using the AKAZE method described previously. A brute-force Knn matcher was used for associating keypoint descriptors, returning only the best match between each pair. A graphical representation of keypoints matching is shown in Fig. 4. The cost function is defined as the sum of all geometric distances between associated keypoints, calculated as a function of the transformation parameters to obtain a global mismatching indicator. The best combination is reached when the cost function reaches its minimum (Eq. 2). Thus, the optimization problem has three dimensions represented
Image Registration Techniques for Independent Acquisitions
275
by the x and y translation and the rotation component. (txbest , tybest , rotbest ) = arg min( tx ,ty ,rot
Nkp
dist(kpref , kpdef (tx , ty , rot)))
(2)
k=1
Table 1. Transformation parameters combinations which define the experiments Exp1
Exp2
Exp3
Exp4
Exp5
Exp6
Exp7
Exp8
tx [px]
10
10
10
10
50
50
50
50
ty [px]
10
10
50
50
10
10
50
50
5
15
5
15
5
15
5
15
Rot [deg]
Fig. 4. Example of keypoint matching between the reference image (left) and the deformed one (right)
The PSO algorithm was initialized with a set of 50 randomly positioned particles within the image boundaries represented by the image domain. These boundaries were enforced also during the optimization process. The maximum iteration number has been set to 1000. However, a tolerance parameter tol, representing the relative error in the objective function, has been introduced for convergence acceptability. In particular, the PSO algorithm stops when the minimum of the cost function is reached, giving as output the optimal transformation parameters. To avoid early stopping, a number of iterations tolIter is defined, in which the optimizer has the opportunity to improve the best cost. For this application, tol = 0.1, tolIter = 50 were chosen. At the end of the experiments, a statistical analysis was performed to analyze the error in parameter estimation across the experiments. Figure 5 represents the relative error with respect to the predicted parameter values as bar plots. As for the cost function, the values have been normalized with respect to the number of keypoints, since the sum of the distances depends on this value. In this case, the relative error is calculated with respect to zero, since the matching keypoints should ideally overlap perfectly. From
276
D. Pennati et al.
these plots it can be seen that in 5 out of 8 experiments the relative error is close to zero for all parameters and there is a strong invariance within each experiment. Also in the other 3 experiments, the relative error is at most 6–7% considering the outliers and less than 5% in the interquartile range. This shows good consistency and robustness of the algorithm not only between different input images but also between different applied deformations. A similar study was performed by Isa-Jara et al. [6] using MRI images and an intensity-based method. In this work, the relative error with respect to the values of the target parameters was about 0.1%, but the analysis was always performed with the same two images. This result is comparable to that obtained with keypoints matching, if outliers are excluded. The greater variability is related to the different image pairs considered in the experiments.
Fig. 5. Statistical error plot of transformation parameters over performed experiments
4 Conclusion The importance of monomodal image registration has been emphasised in this work, together with the need of developing an automatic tool for feature alignment. When dealing with morphological images, rich in anatomical detail, a feature-based approach can be very powerful if a robust keypoints searching and matching algorithm is adopted.
Image Registration Techniques for Independent Acquisitions
277
The AKAZE method has demonstrated its efficiency in computing keypoints thanks to nonlinear diffusion filtering, which allows similar anatomical regions not to be confused. The chosen method has been shown to be invariant to rototranslations. As for the optimization algorithm, the PSO method provided consistent results both when varying the input image, and when changing the initial deformation matrix. A relative error of 5% or less is a consistent result because this discrepancy is close to DICOM resolution (up to 0.1 mm in the considered acquisition equipment) and for this reason inappreciable. The correct initialization of the algorithm is crucial, since the particles must be able to move freely, but without exceeding the limits of the domain range. The main bottleneck of PSO and of genetic algorithms in general is the speed: this has been increased by introducing the parameters tol and tolIter parameters, but the optimization cycle is still too slow for a real-time application. This aspect can be solved by parallelizing the particle evaluation or by reducing the number of keypoints in regions of interest where clustering is present.
References 1. Abdel-Basset, M., Fakhry, A.E., El-Henawy, I., Qiu, T., Sangaiah, A.K.: Feature and intensity based medical image registration using particle swarm optimization. J. Med. Syst. 41(12), 1–15 (2017) 2. Alcantarilla, P.F., Solutions, T.: Fast explicit diffusion for accelerated features in nonlinear scale spaces. IEEE Trans. Patt. Anal. Mach. Intell 34(7), 1281–1298 (2011) 3. Ballerini, L.: Particle swarm optimization in 3d medical image registration: a systematic review (2023). arXiv preprint arXiv:2302.11627 4. El-Gamal, F.E.Z.A., Elmogy, M.M., Atwan, A.: Current trends in medical image registration and fusion. Egypt. Inform. J. 17, 99–124 (2016) 5. Gupta, S., Gupta, P., Verma, V.S.: Study on anatomical and functional medical image registration methods. Neurocomputing 452, 534–548 (2021) 6. Isa Jara, R., Buchelly, F.J., Meschino, G.J., VL, B.: Improved particle swarm optimization algorithm applied to rigid registration in medical images. In: VII Latin American Congress on Biomedical Engineering CLAIB 2016, Bucaramanga, Santander, Colombia, October 26th– 28th, 2016, pp. 161–164. Springer (2017) 7. John, L., John, L.: A review of image registration methods in medical imaging. Int. J. Comput. Appl. 178, 38–45 (2019). https://doi.org/10.5120/ijca2019918884 8. Kutaish, H., Acker, A., Drittenbass, L., Stern, R., Assal, M.: Computer-assisted surgery and navigation in foot and ankle: state of the art and fields of application. EFORT Open Rev. 6(7), 531–538 (2021) 9. Mani, V., Arivazhagan, S.: Survey of medical image registration. J. Biomed. Eng. Technol. 1(2), 8–25 (2013) 10. Misra, I., Rohil, M.K., Manthira Moorthi, S., Dhar, D.: Feature based remote sensing image registration techniques: a comprehensive and comparative review. Int. J. Remote Sens. 43(12), 4477–4516 (2022) 11. Tareen, S.A.K., Saleem, Z.: A comparative analysis of sift, surf, kaze, akaze, orb, and brisk. In: 2018 International Conference on Computing, Mathematics and Engineering Technologies (iCoMET), pp. 1–10. IEEE (2018)
Extremity Bones Segmentation in Cone Beam Computed Tomography, a Novel Approach Eleonora Tiribilli1,2(B) , Leonardo Manetti2 , Leonardo Bocchi1 , and Ernesto Iadanza1,3,4 1 Department of Information Engineering, University of Florence, Florence, Italy
[email protected]
2 Epica Imaginalis, Sesto Fiorentino, Italy 3 Department of Medical Biotechnologies, University of Siena, Siena, Italy 4 IFMBE Council of Societies Chair, Florence, Italy
Abstract. Bone segmentation in Cone Beam Computed Tomography (CBCT) may appear as a simple and efficient task. Unfortunately, this is not true for complex areas such as the hand, wrist or feet, where there are many small and thin bones that are very difficult to distinguish. In this article an efficient graph cut method to segment complex districts in CBCT has been developed. A graph cut approach is initialized with an automatic “background” and “object” evaluation, carried out with pixel-based techniques, such as Otsu and Yen thresholding. Segmentation result is post-processed with morphological operation, then connected components with 26-connectivity are evaluated to separate the different bone. Finally, the user can isolate the bone (or bones) of interest and display a 3D model. The approach is compared with a standard graph cut approach whit requires user scribbles to segment a single bone, both in terms of accuracy and usability. Keywords: Segmentation · Computed tomography · Bones · Foot and ankle · Joints · Hand and wrist · Graph cut · Connected component · User interaction
1 Introduction Cone-beam computed tomography (CBCT), an established imaging modality, is discovering new uses in the orthopedic field. The main benefit of CBCT scanners over spiral computed tomography is the capability to obtain 3D images more quickly, accurately, and with less radiation exposure, using smaller imaging devices [1]. That last point has made CBCT an effective imaging technique in emergency department and surgical rooms, specially for the diagnosis and treatment planning in extremities [2]. Clinical applications of this imaging modality will benefit from the development of more complex image analysis tools. In this sense, 3D models of bones are of great interest in the orthopedic field, for intra operative guidance, reconstructive surgery and implant design. Moreover, new technologies such as augmented reality or 3D printing rely on the accuracy of 3D models. An accurate bone segmentation of CBCT volume is a crucial step to create three-dimensional models for orthopedics field. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 278–284, 2024. https://doi.org/10.1007/978-3-031-49062-0_30
Extremity Bones Segmentation in Cone Beam Computed Tomography
279
A long bone’s center section, also known as the shaft or diaphyses, is made up of hard, strongly attenuating, and thick cortical tissue that has higher values on tomographic scans. Segmenting the shaft of a long bone is a reasonably easy task because of the smooth bone borders and strong contrast between the cortical layer and the surrounding low-intensity soft tissue. Global thresholding, followed by fundamental morphological operations, can produce effective outcomes [3]. In contrast, short bones (cube-shaped with roughly equal vertical and horizontal dimensions) are made of spongy bone, covered with a very thin layer of compact bone. The wrist and ankle bones are examples of this kind of hard tissues. As Sebastian et al. [4] listed, there are manifold challenges regarding the segmentation of these kind of hard (shorts bones and epiphyses) tissues from tomography volumes: Low-contrast and weak bone boundaries: thickness of the cortical shell is lower than in long bones. Varying density of cancellous tissues: cancellous bone tissue is vascular and spongy, giving the appearance of textured and inhomogeneous bone. Thus, bone tissues density characteristics cannot be uniformly described. Hence, simple segmentation methods such as region growing or thresholding unavoidably produce inadequate results [4]. Narrow inter-bone spacing: the articular space between adjacent bones is extremely narrow. The inter-bone zones are diffused and brighter due to the partial volumetric effect, which is a characteristic of the CT modality and makes the bones appear to be in close proximity. Bone segmentation in the extremities can be recognized as a challenging task. Moreover, extremities such as foot, ankle, hand and wrist, are characterized by a huge number of small, asymmetrical-shaped structures and densities. Image processing solutions must be designed to facilitate and expedite this work while minimizing manual interaction and inter-operator variability [5]. The application of graph cut for user-friendly single bone segmentation and labeling in CBCT of extremities is discussed in this work. Graph cut is a well known approach, Boykov et al. [6] applied it to image segmentation for the first time. In the past years, many researchers have applied this method to bone segmentation. Aslan et al. [7] use this approach to segment a single vertebra in CBCT scan. Liu et al. [8] interactively separate bones in CT scans, applying graph cut to segmented images. Pauchard et al. [9] use graph cut to compute a femur finite element model from clinical computed tomography scan for hip fracture prediction. In conventional graph cut algorithm, the user makes scribbles on the input image to select background and object [10]. In this work an automatically seeded algorithm is presented, in which the initialization has been done with a mask obtained through automatic thresholding. This allows to have a larger number of seeds pixels and results in an improved accuracy of segmentation. Moreover, this novel approach reduces user interaction to segment a single bone, as no background scribble from the user is required. Once all the bones were segmented, they are labelled via a connected component algorithm, then the bones of interest are extracted through a user interaction. In this article, the overall approach is illustrated and some results are reported. Then, these results are compared with the traditional graph cut approach in the aim of extract a single bone, in term of accuracy and usability.
280
E. Tiribilli et al.
2 Materials and Methods The aim of this work is to segment a single bone in Cone Beam-CT scan of the extremities, in an accurate and user-friendly manner. To develop and test our method ten Cone Beam CT scans of the extremities were available. In this section, a segmentation algorithm based on graph cut approach is presented. Graph cut is a segmentation approach that considers both regional and boundary characteristics of the volume, this makes it suitable for segmentation in our target area. In graph cut segmentation the volume is represented as a graph [11, 12], a graph G = (V, E) is a generic structure made up of a collection of nodes V that represent the original image’s pixels or voxels and a set of arcs (or edges) E that connect the nodes. Two unique terminal nodes, source s and sink t, which stand in for “object” and “background,” respectively, in biobject segmentation are present in addition to nodes set V. As shown in Fig. 1 the graph has two different kinds of edges represented in yellow and in blue or red. The first form, called n-links, which connects nearby pixels, uses the prefix “n” for “neighbor.” The second kind of t-links, where “t” stands for “terminal,” joins terminals and pixels. In a graph G, a s/t cut is the split of V into two disjoint subsets S and T, with all object voxels connected to an object terminal node S and all background voxels attached to a background terminal node T. The objective is to choose the appropriate cut that will produce the best results given the segmentation requirements. The cheapest cut is also the best cut because it is the least expensive. The following energy function, which takes into account both regional and boundary factors, can be minimized to get the best cut. Rp (Ap ) + Bp,q (Ap , Aq ) (1) E(A) = p∈P
(p,q)∈N
Ap is the label of pixel p in an image P. The first term is a regional term and Rp (Ap ) is the penalty to assign label Ap to pixel p. The region energy function should reach the smallest value if labels are correctly assigned to all the pixels. The second term is a boundary term Bp,q (Ap , Aq ), which can be interpreted as a penalty for discontinuity between p and q. Bp,q (Ap , Aq ) is large when p and q are similar and near to zero when p and q are completely different. In another word, if p and q are similar, then the probability that they belong to the same object is high. Otherwise, p and q may belong to different objects. Therefore, boundary energy is small if neighboring pixels p and q are different.
Fig. 1. Graph cut segmentation algorithm.
Extremity Bones Segmentation in Cone Beam Computed Tomography
281
N-links (neighbor-links) connect (typically) two neighboring nodes and are associated with the boundary terms. In our implementation their cost is wp,q = f(|Ip − Iq |)
(2)
where Ip and Iq are intensities at pixels p and q and f(x) = K exp(−x2 /σ2 ) T-links (terminal-links) represent regional image properties and are connected to two terminals in the graph. For the smallest cost cut, inexpensive edges are attractive choices. The object and background model determines the weights of the T-links Rp to the object and the background vertex. Rp (obj) = −ln(Pr < Ip |O >))
(3)
Rp (bkg) = −ln(Pr < Ip |B >))
(4)
where Pr and Pr represent likelihoods for object and background and are given by Gaussian mixture model. These likelihoods for background and object are constructed based on an initialization mask, in which each pixel is marked as background object or probable background or object. Accordingly with the characteristic of extremity bones in CBCT, an automatic labeling of volume data is performed. This is achieved using the well-known threshold techniques Yen and Otsu [13, 14]. With Yen algorithm, a threshold to identify sure foreground pixels is calculated. Otsu method determines a threshold to identify sure background in the volume. All the pixels included in the range between these thresholds are labeled as uncertain pixels to be evaluated from the graph cut algorithm. The segmented volume resulting from this algorithm is post processed with morphological operations. Bones making up extremities are separated via connected components labeling approach in three dimensions using a 26-connected neighborhood. Ultimately, a Graphic User Interface (GUI) developed in Python allows the user to select the bone of interest. The segmented bones are rendered in 3D.
3 Results In this work, a dataset of ten CBCT volumes is evaluated. The dataset consists of four feet scans, four wrist scans and two knee acquisitions. In Figs. 2, 3 and 4 the performances of the presented approach on foot and wrist are presented. The algorithm is capable to segment and labels all the bones in these districts and the user can easily select the bone of interest, quickly switching among adjacent bones. The benchmark for testing our approach performances is the classical Graph Cut approach in which user makes a scribble to initialize the algorithm and to calculate object and foreground likelihood. In term of usability, our approach requires less user interaction and provides more flexibility in segmenting adjacent bones. Because the user select bones after the entire
282
E. Tiribilli et al.
Table 1. Comparison of the results on wrist and foot dataset. Dice Coefficient Score (DSC) is provided for both the proposed Graph Cut (GC) method and Graph Cut (GC) with user scribbles Target bones
GC scribbles
GC threshold
Trapezium
0.867
0.896
Scafoid
0.890
0.901
Lunate
0.823
0.896
Navicular
0.882
0.897
Talus
0.875
0.891
Medial cuneiform
0.793
0.820
Proximal falangis
0.784
0.882
Navicular
0.896
0.898
district is segmented, this allows a faster selection of the desired bone. In quantitative terms, the well know metrics of Dice Score Coefficient (DSC) is used. A manual segmentation of specific wrist and foot bones is performed using 3D Slicer and then used as the ground truth for metrics evaluation. Table 1 reports performance in term of DSC between our approach and classical graph cut, in the segmentation of specific bone targets in wrist and foot.
(a) Graph cut segmentation and con- nected component labeling
(b) Capitate bone selection with a click on the corresponding label
Fig. 2. Wrist bones segmentation obtained with the proposed approach.
Extremity Bones Segmentation in Cone Beam Computed Tomography
(a) Graph cut segmentation and con- nected component labeling
283
(b) Navicular bone selection with a click on the corresponding label
Fig. 3. Foot bones segmentation obtained with the proposed approach.
Fig. 4. 3D rendering of the Navicualr bone obtained with the proposed approach
4 Conclusions CBCT imaging modality is discovering new applications in orthopedics, especially in the extremities. In this work, a novel approach to fast and accurate segment single bone in the extremities is presented. Extremities 3D volume data obtained by the promising imaging techniques of CBCT has been used. First, the algorithm labels volume according to Yen and Otsu thresholding in order to create a map of pixel used to initialize Graph Cut. Then, the algorithm computes bone segmentation. Connected components are computed to label different bones. Finally, the user selects a single bone of interest (or several ones). A 3D rendering of bone is proposed. This method has been compared with a standard graph cut approach in which the user has to labeling manually the bone of interest as “object” and the others as “background”.
284
E. Tiribilli et al.
Results demonstrate that our approach is more accurate, since it’s better in term of the dice coefficient score. Regarding system usability allows less user interaction, resulting more user-friendly.
References 1. Bailey, J., Solan, M., Moore, E.: Cone-beam computed tomography in orthopaedics. Orthop. Trauma 36(4), 194–201 (2022) 2. Jacques, T., Morel, V., Dartus, J., Badr, S., Demondion, X., Cotten, A.: Impact of introducing extremity cone-beam ct in an emergency radiology department: a population-based study. Orthop. Traumatol. Surg. Res. 107(2), 102834 (2021) 3. Haralick, R., Sternberg, S., Zhuang, X.: Image analysis using mathematical morphology. IEEE Trans. Pattern Anal. Mach. Intell. 36(9), 532–520 (1987) 4. Sebastian, T., Tek, H., Crisco, J., Kimia., B.: Segmentation of carpal bones from ct images using skeletally coupled deformable models. Med. Image Anal. 7(9), 21–24 (2003) 5. Tiribilli, E., Bocchi, L., Iadanza, E.: Bones segmentation techniques in computed tomography, a survey. In: Proceedings IUPESM World Congress on Medical Physics and Biomedical Engineering (2022) 6. Boykov, Y., Jolly, M.P.: Interactive graph cuts for optimal boundary & region segmentation of objects in n-d images. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, vol. 1, pp. 105–112 (2001) 7. Aslan, M., Ali, A., Chen, D., Arnold, B., Farag, A., Xiang, P.: 3d vertebrae segmentation using graph cuts with shape prior constraints. In: ICIP International Conference on Image Processing, pp. 2193–2196 (2010) 8. Liu, L., et al.: Interactive separation of segmented bones in ct volumes using graph cut. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2008: 11th International Conference, New York, NY, USA, September 6–10, 2008, Proceedings, Part I 11, pp. 296–304. Springer (2008) 9. Pauchard, Y., et al.: Interactive graph-cut segmentation for fast creation of finite element models from clinical ct data for hip fracture prediction. Comput. Methods Biomech. Biomed. Eng. 19(16), 1693–1703 (2016) 10. Yi, F., Moon, I.: Image segmentation: a survey of graph-cut methods. In: 2012 International Conference on Systems and Informatics (ICSAI2012), pp. 1936–1941 (2012) 11. Jirik, M., Zelezny, M.: Image segmentation in medical imaging via graph-cuts. In: 11th International Conference on Pattern Recognition and Image Analysis: New Information Technologies (PRIA-11–2013). Samara, Conference Proceedings (2013) 12. Boykov, Y., Funka-Lea, G.: Graph cuts and efficient N-D image segmentation. Int. J. Comput. Vis. 70(2), 109–131 (2006) 13. Yen, J.C., Chang, F.J., Chang, S.: A new criterion for automatic multilevel thresholding. IEEE Trans. Image Process. Publ. IEEE Signal Process. Soc. 4(3), 370–378 (1995) 14. Otsu, N.: A threshold selection method from gray-level histograms. IEEE Trans. Syst. Man Cybern. 9(1), 62–66 (1979)
Deep-Learning Based Automatic Determination of Cardiac Planes in Survey MRI Data Jan Jurca1 , Vratislav Harabis1(B) , Roman Jakubicek1 , Tomas Holecek1,2 , Petra Nemcekova1 , Petr Ourednicek2 , and Jiri Chmelik1 1 Department of Biomedical Engineering, FEEC, Brno University of Technology, 3082/12, 616
00 Brno, Czech Republic [email protected], [email protected] 2 Department of Imaging Methods, St. Anne’s University Hospital Brno, Pekarska 664/53, 656 91 Brno, Czech Republic
Abstract. Inference of the radiological planes of the heart in MRI is a crucial step for valid data acquisition to examine the structure and function of the human heart in detail. In this paper, we present a deep learning model for automatic inference of the radiological plane of the heart from 3D survey sequences. The proposed neural network is based on the V-Net [6] architecture that has been developed to perform inference on the radiological positions of the hearts. The network is designed to take a 3D image as input and generate a regressed heatmap of probable plane positions as output. The results show that the proposed method is feasible for automatic geometry planning. It has the potential to increase the efficiency of medical imaging. The presented networks show that they can locate cardiac landmarks even from data with anisotropic voxels. It can improve the accuracy and speed of diagnosis, allowing for faster and more effective treatment. Keywords: Heart axis determination · Regression · Deep-learning · MRI
1 Introduction Medical imaging plays a vital role in the diagnosis and treatment of various dis- eases, including heart diseases. The radiological planes of the heart are used to examine the structure and function of the heart in detail. These planes are usu- ally acquired through various imaging techniques such as Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) see [4]. However, identifying the correct radiological orientation requires specialised medical knowledge and experience. 1.1 Motivation Correct radiological orientation is crucial for the accurate interpretation of heart medical images. When the heart is viewed in the correct orientation, it is easier to identify specific structures and diagnose various cardiac pathologies. For example, the four-chamber view is a crucial view for assessing the function of the heart chambers and valves. When © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 285–292, 2024. https://doi.org/10.1007/978-3-031-49062-0_31
286
J. Jurca et al.
the heart is not orientated correctly, it can lead to misinterpretation of the image and misdiagnosis. Automatic determination of the heart plane from 3D data is an important step toward ensuring that medical images are viewed in the correct orientation, which can lead to more accurate diagnoses and better patient outcomes. 1.2 State-Of-The-Art Methods Today, in existing medical systems, the radiological planes of the heart are still determined manually by the radiologist, but there are helpful approaches to facilitate data navigation and scan planning. However, the time saved during planning using these algorithms is quite minor [2] for the experienced radiologist. The authors of [3] proposed the method for automatic scanning planning. This method is based on a model-based segmentation approach, where a triangulated surface model [6]) composed of the seven main parts was used. In this method, the affine transformation is used first for pose optimisation using a global similarity transformation. In the second step, the energy-minimising free-form deformation using sequencespecially trained, locally varying boundary descriptors is used for precise model estimation. The results show that the proposed method is feasible for automatic geometry planning with an overall success rate of 94%. The limitation of this study is the relatively low number of patients tested (50 patients). Additionally, testing is lacking in a wide spectrum of various cardiovascular pathologies. A model-based method for automatic view planning for cardiac MRI acquisition was proposed in [5]. In this paper, the left ventricle (LV) mesh representation is created. To estimate the LV pose and extract LV boundaries, probabilistic boost trees are used. The proposed method was tested using 173 localiser volumes (3D full chest magnetic resonance parallel imaging within a single breath hold) from 100 subjects. The result shows that the method is capable of view planning, but the authors also conclude that low image quality and large image variation (pathologies, etc.) are challenging. The authors in [1] propose a method based on deep learning to estimation MRI planes. It is a retrospective study in which the authors used 482 cardiac magnetic resonance studies. Studies contain 892 long-axis cone steady-state free precession series. The authors implemented 2D U-Net and modified it to heatmap regression. They used this network for the detection of landmarks (mitral valve and heart apex). 2.5D U-Net was used to create a bounding box around the heart to reduce the search space for anatomic landmarks. The results show that the deep learning-based method is capable of localising cardiac landmarks for the prescription of the four-chamber plane (4CH), three-chamber plane, twochamber plane (2CH) and short-axis view (SA). However, the authors stated that it is a proof-of-concept feasibility study. It is not clear whether this method obtains valid results with data obtained using various MRI systems, especially in various survey sequences (lower image quality, artefacts, etc.), which are used to determine planes. The regression network model was also used by the authors in [2] to determine the heart planes from the scout volume data. In the first step, they used the Otsu method to segment the heart regions and then used the trained regression network to calculate the distance map to the plane. Finally, they use the leastsquares method to fit the valid plane position. They also proposed a shimming method based on the V-Net structure. This network is trained to separate heart and nonheart regions, and final masks are used to
Deep-Learning Based Automatic Determination of Cardiac Planes in Survey MRI Data
287
calculate shimming currents for magnetic field optimisation. The results show that the deep learning-based method can minimise operator dependence and can reduce overall scanning time by 13%. The study still has limitations. The proposed method was tested on quite a low number of data, where the cohort contains only 10 healthy volunteers and 10 cancer patients, and it is unclear whether this method is independent of anatomical variabilities, some artefacts, various heart diseases, etc. 1.3 Presented Work In this paper, we present a deep learning model for automatic inference of the radiological planes of the heart. The neural network that we propose is based on the V-Net [6] architecture that has been developed to perform an estimation on the radiological planes of the heart. The proposed neural network is designed to take a 3D image of the heart as input and generate a regressed heatmap of probable plane positions as output, and from the heatmap can establish the angle of the plane. This approach minimises the need for specialised medical knowledge and expertise, making it easier for medical professionals to diagnose and treat heart diseases. The use of a neural network to generate regressed heat maps of probable plane positions can improve the accuracy and speed of diagnosis, allowing faster and more effective treatment. It has the potential to increase the efficiency of medical imaging-based diagnosis.
2 Methodology 2.1 Data Description For this purpose, the publicly available dataset [7] contains 30 cardiac magnetic resonance scans. MRI acquisition was performed on a 1.5 T Philips Achieva scanner using a 3D balanced steady-state free precession acquisition. Data were acquired from a whole heart in end-diastole gated with ECG and in free-breathing mode with respiratory gating and with voxel resolution of 1.25 × 1.25 × 2.7 mm3 . 2.2 Manual Cardiac Planes Determination Therefore, the data set is primarily aimed at left atrial segmentation, ground truth (i.e. parameters of the long axis (LA) vector, SA view, 4CH and 2CH plane) had to be obtained manually by the radiologist expert in each scan. For the data annotation, we have prepared a custom graphical interface that allowed us to determine planes (the ground truth through) in 3D space. The LA vector is defined as two points that represent the mitral valve and the heart apex in the 3D space. The format of ground truth data is a 3D matrix, which corresponds to dimensions of the annotated scan, where the values in the matrix are in the range of 0 to 1 and represent the probability that the particular plane intersects each position. We can describe the annotated matrix as a radiological plane heatmap, the higher the value, the greater the probability of plane intersection at given position. Visualisation of the annotated data is shown in Fig. 1.
288
J. Jurca et al.
Fig. 1. Illustration of the annotations of the data. Each matrix represents one radiological plane.
2.3 CNN-Based Cardiac Planes Determination For task of determination of the heart planes we used the standard image to image convolutional network V-Net [6]. The input of the network is the entire scan in standard anatomical orientation, and the output is the heat map of the particular radiological plane. Several transformations are performed before the input data are transmitted to the network. First, the scan has to be resized and represented as a 3D matrix of dimension 128 128 128. The values of the matrix are then normalised to an interval from 0 to 1. These data are passed on to the network, and its output is in the form of heat maps that represent the probability of the position of the plane. There are basically two variants of networks according to the output. The first is that one network is dedicated to the determination of one specific plane (e.g. 2CH). This means that the output of this network is only one channel 3D matrix (as we call it a Single-shot network). Or one network can determine all planes at once (as we call it a Multi-shot network). This means that this network has multichannel output, where each channel is interpreted as a determined heatmap for a defined heart plane. Training of both variants is very similar. Training data preparations. As the size of the dataset and the resulting amount of training data are small, we had to use the data augmentation approach to improve the network’s ability to generalise over different data. Augmentation consisted of two operations. The first is the random rotation of 45z on each side around each axis. Second, we applied random additive Gaussian noise to the input image. For each image from the dataset, 20 other images were generated using these operations. The data set was split into two parts. The training part has 22 (440 with augmented) samples and the testing part has 8 (160 with augmented) remaining samples. Training. V-Net regression network was trained for 30 epochs in the described dataset using Adam optimiser and Mean Squared Error (MSE) as loss function. We used the following setting of hyperparameters: initial learning rate 0.001, β 1 = 0.9, β 2 = 0.999, batch size 2, weight decay 1e-8. Exponential Linear Unit (ELU) activation functions with α = 1.0 were used in all convolution layers together with 3D batch
Deep-Learning Based Automatic Determination of Cardiac Planes in Survey MRI Data
289
normalisation. In the two deepest downconvolution and up-convolution layers, a 3D DropOut regularisation was used with p equal to 0.5. Source code, annotated data, and trained networks are available from the GitHub repository https://github.com/janjurca/HeartNet.
3 Results and Discussion The proposed networks were tested with data from eight patients (one volumetric image for each patient). The error angle is computed as the angle between the estimated plane and the annotated plane in degrees. The error of the estimated angles is shown in Fig. 2 for Single-shot on the left and Multi-shot on the right. It shows that the Multi-shot network is able to estimate planes with slightly lower error than the Single-shot network, especially for the 4CH plane (mean error for Multi-shot is 5.03z ; mean error for Singleshot is 7.01z ). Also, the Multi-shot network with a mean error 6.86° estimate SA view slightly better than the Single-shot network with a mean error angle 6.97z . Estimation of the 2CH plane is quite stable for both networks, and the single-shot network has a lower mean error 5.92z than the Multi-shot network with a mean error of 6.25z . That means the Multi-shot network is more suitable because training is less time-consuming, uses fewer computing resources, and results are quite similar or better than the Single-shot network.
Fig. 2. The error in the estimated angle of the planes (2CH, 4CH, and SA view) in degrees. The red line represents the median, the blue box represents the 25 to 75 percentile, and the black whiskers represent the maximum and minimum.
Figure 3 shows two examples of the correct estimation of the SA view, the 2CH and 4CH planes. As for the SA view, it is correctly tilted in both cases in the result intersecting the left and right ventricles. The circular shape of the myocardium indicates the correct inclination of the plane whose normal vector is in the direction of the long axis of the heart. In planes 2CH, the left atrium is clearly visible in its entire cross section with the atrial appendage and mitral valve. The four chamber planes show correctly both atrium, tricuspidal and mitral valves for possible comparison of atrium size or valve
290
J. Jurca et al.
Fig. 3. Examples of all three automatically estimated planes for two cases.
functionality. Both the ventricles and the myocardium are also distinct in these planes to assess possible morphologies. Two selected examples of the estimated planes to compare the single and multiple shot prediction approaches separately for 4CH and 2CH can be seen in Fig. 4. In the case of 4CH estimation, a plane that intersects the ascending aorta is predicted by the Singleshot network, which is not desirable for 4CH (in-dicated by the blue arrow). For the 2CH plane, the Multi-shot approach suggests a better inclination of the predicted plane, particularly in the left atrium area (indicated by the green arrow), which is visible across the entire cross-section. Here is also the more prominent left atrial appendage marked with a red arrow. The single-shot network proposes a slightly different plane inclination, resulting in a visibly smaller left atrial cross section, which may seem correct. However, comparing it with the Multi-shot network’s estimate shows that the inclination can still be optimised. The provided sample examples graphically confirm that the Multi-Shot approach yields slightly better results, which also correlates with the objective evaluation.
Deep-Learning Based Automatic Determination of Cardiac Planes in Survey MRI Data
291
Fig. 4. Examples of estimated planes for comparison of Multi-shot and Single-shot network. Blue arrow shows incipient ascending aorta; red arrow shows the left atrial appendage; green arrow points to the larger left atrium.
4 Conclusion The proposed and presented networks show that they are capable of estimating radiological planes even from data with anisotropic voxels. The main limitation is the stil low number of images in the public dataset and also the apex of the heart missing in some images. This means that the annotations of the data were quite challenging. However, this public dataset was used for the feasibility study of this deep-learning approach to perform inference on the radiological planes of the heart from 3D survey sequences, which have a resolution and image quality similar to that of the dataset used. As part of further research, we have started collecting cardiac magnetic resonance data that contain survey sequences (standard and also research oriented high-resolution surveys). These data also contain radiological planes defined by a radiology specialist. This will make it possible to expand the dataset and to allow the networks to be re-trained, which increases its robustness and applicability. We also wanted to compare our method with the published method in [2], but the authors told us that they could not provide their data and learned network for comparison. Acknowledgment. This paper and the research behind it would not have been possible without the support of Philips Healthcare company. Computational resources were provided by the Ministry of Education, Youth and Sports of the Czech Republic under the CESNET (Project No. LM2015042) and CERIT-Scientific Cloud (Project No. LM2015085) provided within the programme Projects of Large Research, Development and Innovations Infrastructures. The Titan Xp GPU used for the implementation of the algorithms was donated by NVIDIA Corporation to support academic research.
292
J. Jurca et al.
References 1. Blansit, K., Retson, T., Masutani, E., Bahrami, N., Hsiao, A.: Deep learning–based prescription of cardiac mri planes. Radiol. Artif. Intell. 1(6), e180069 (2019), https://doi.org/10.1148/ryai. 2019180069, pMID: 32090204 2. Edalati, M., Zheng, Y., Watkins, M.P., Chen, J., Liu, L., Zhang, S., Song, Y., Soleymani, S., Lenihan, D.J., Lanza, G.M.: Implementation and prospective clinical validation of ai-based planning and shimming techniques in cardiac mri. Med. Phys. 49(1), 129–143 (2022). https:// aapm.onlinelibrary.wiley.com/doi/abs/10.1002/mp.15327 3. Frick, M., Paetsch, I., den Harder, C., Kouwenhoven, M., Heese, H., Dries, S., Schnackenburg, B., de Kok, W., Gebker, R., Fleck, E., Manka, R., Jahnke, C.: Fully automatic geometry planning for cardiac mr imaging and reproducibility of functional cardiac parameters. J. Magn. Reson. Imag. 34(2), 457–467 (2011). https://onlinelibrary.wiley.com/doi/abs/https://doi.org/10.1002/ jmri.22626 4. Kramer, C.M., Barkhausen, J., Bucciarelli-Ducci, C., Flamm, S.D., Kim, R.J., Nagel, E.: Standardized cardiovascular magnetic resonance imaging (cmr) protocols: 2020 update. J. Cardiovasc. Magn. Reson. 22(1), 17 (2020). https://doi.org/10.1186/s12968-020-00607-1 5. Lu, X., Jolly, M.P., Georgescu, B., Hayes, C., Speier, P., Schmidt, M., Bi, X., Kroeker, R., Comaniciu, D., Kellman, P., Mueller, E., Guehring, J.: Automatic view planning for cardiac mri acquisition. In: Fichtinger, G., Martel, A., Peters, T. (eds.) Medical image computing and computer-assisted intervention – MIC- CAI 2011. pp. 479–486. Springer Berlin Heidelberg, Berlin, Heidelberg (2011). https://doi.org/10.1007/978-3-642-23626-6 6. Milletar‘ı, F., Navab, N., Ahmadi, S. A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. 2016 Fourth International Conference on 3D Vision (3DV), pp. 565–571 (2016), https://doi.org/10.1109/3DV.2016.79 7. Peters, J., Ecabert, O., Meyer, C., Schramm, H., Kneser, R., Groth, A., Weese, J.: Automatic whole heart segmentation in static magnetic resonance image volumes. In: Ayache, N., Ourselin, S., Maeder, A. (eds.) Medical Image Computing and Computer-Assisted Intervention—MICCAI 2007. pp. 402–410. Springer Berlin Heidelberg, Berlin, Heidelberg (2007). https://doi.org/10.1007/978-3-540-75759-7 8. Tobon-Gomez, C., Geers, A.J., Peters, J., Weese, J., Pinto, K., Karim, R., Ammar, M., Daoudi, A., Margeta, J., Sandoval, Z., Stender, B., Zheng, Y., Zuluaga, M.A., Betancur, J., Ayache, N., Chikh, M.A., Dillenseger, J.L., Kelm, B.M., Mahmoudi, S., Ourselin, S., Schlaefer, A., Schaeffter, T., Razavi, R., Rhode, K.S.: Benchmark for algorithms segmenting the left atrium from 3D CT and MRI datasets. IEEE Trans. Med. Imag. 34(7), 1460–1473 (2015). https://doi. org/10.1109/TMI.2015.2398818
‘3D Printed Breast Phantoms Materials for X-ray Imaging Techniques’ Aris Dermitzakis1(B) , Martin Pichotka2 , Antzela Petrai1 , Moritz Weigt2 , and Nicolas Pallikarakis1 1 Biomedical Technology Unit, Department of Medical Physics, Faculty of Medicine,
University of Patras, Patras, Greece [email protected] 2 Division of Medical Physics, Department of Diagnostic and Interventional Radiology, University Medical Center Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany
Abstract. Physical breast phantoms are crucial for the development, evaluation, and optimization of X-ray breast imaging systems, such as mammography, digital breast tomosynthesis, and breast computed tomography. Traditional manufacturing methods can be time-consuming and lack precision. This study investigates the use of thermoplastic filaments and photopolymer resins as 3D printing materials for fabricating physical breast phantoms. The linear attenuation coefficients (µ) of these materials were determined to establish a standard attenuation coefficient graph for future research in X-ray imaging. Fused Deposition Modeling (FDM) and Stereolithography (SLA) techniques were employed for 3D printing, using eight different materials. X-ray imaging experiments were performed using a spectral photon-counting µCT-scanner. Preliminary results indicate a mismatch between theoretical and measured spectral attenuation profiles, a number of possible reasons are discussed. Future research will explore alternative equipment and techniques to address these differences. Keywords: Breast phantoms · 3D printing · X-ray imaging
1 Introduction Physical breast phantoms play a critical role in the development, evaluation, and optimization of X-ray breast imaging systems, such as mammography, digital breast tomosynthesis, and breast computed tomography [1]. Breast phantoms are designed to mimic the radiographic properties of real breast tissue and are used for various purposes, including the assessment of image quality, the calibration of imaging systems, and the evaluation of dose optimization algorithms [2]. Furthermore, breast phantoms are essential for training and education purposes, allowing medical professionals to improve their skills in interpreting and diagnosing breast-related pathologies. One of the challenges in creating breast phantoms is to accurately represent the physical properties of breast tissue, including its composition, structure, and X-ray attenuation © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 293–300, 2024. https://doi.org/10.1007/978-3-031-49062-0_32
294
A. Dermitzakis et al.
properties [3]. Traditional methods for manufacturing breast phantoms involve the use of tissue-equivalent materials, such as wax and gelatin, which can be time-consuming, labor-intensive, and may lack precision [4]. Recent advancements in 3D printing technology have provided an alternative approach for the fabrication of breast phantoms. 3D printing, also known as additive manufacturing, enables the creation of complex, customized structures with high precision and reproducibility [5]. Various polymers can be used as printing materials to mimic the different types of breast tissue, such as glandular and adipose tissue, by adjusting their X-ray attenuation properties. The use of 3D printing techniques, such as Fused Deposition Modeling (FDM) and Stereolithography (SLA), has the potential to revolutionize the fabrication of breast phantoms, allowing for greater flexibility, accuracy, and efficiency in the development and assessment of X-ray breast imaging systems [6]. The purpose of this study is to investigate the use of various thermoplastic filaments and photopolymer resins as 3D printing materials for the fabrication of physical breast phantoms. By determining the linear attenuation coefficients (µ) of these materials, we aim to establish a standard attenuation coefficient (µ) graph that can be used for future research and demonstration of breast tissue in X-ray imaging. The printing procedures took place in the Medical Physics Laboratory, which is accommodated in the Department of Medicine—Clinical Laboratory Building at the University of Patras, and the X-ray imaging examinations were performed at AMIR Preclinical Research Center, University Medical Center Freiburg, Germany [7, 8]. 1.1 Equipment and Materials In this study, eight materials were selected and investigated for their potential as physical breast phantoms in X-ray breast imaging. These materials were chosen based on their ability to mimic the radiographic properties of breast and adipose tissue. The materials consist of six thermoplastic filaments, namely ABS, PLA, PLA_Pro, CPE, Nylon, and PET_G, and two photopolymer resins, Clear and Purple as seen in Fig. 1. The thermoplastic filaments were printed using Fused Deposition Modeling (FDM) technology with the Prusa i3 MK3S printer, while the photopolymer resins were printed using Stereolithography (SLA) technique with the Nobel Superfine printer [9–11]. Following the printing process, the photopolymer resins required UV curing in a specialized chamber to solidify and stabilize their structure. The samples were printed in the form of 3D cubes with dimensions of 20 × 20 × 20 mm to ensure uniformity and comparability across the different materials. 1.2 Fused Deposition Modeling Settings (FDM) To ensure optimal printing conditions for each thermoplastic filament, specific settings were adjusted during the FDM process. These settings include fill density, fill pattern, bottom-top fill pattern, fill angle, layer height, extrusion multiplier, nozzle temperature, bed temperature, print speed perimeters, and fan speed (Table 1).
‘3D Printed Breast Phantoms Materials for X-ray Imaging Techniques’
295
Fig. 1. The 3D printed cubes of Purple, Clear, CPE, Nylon, PET-G, PLA_Pro, PLA, ABS (from left to right) Table 1. Settings of the thermoplastics filaments for 3D printing technique Filaments
ABS
PLA
PLA_Pro
Nylon
CPE
PET_G
Fill Density 100%
100%
100%
100%
100%
100%
Fill pattern
Rectilinear
Rectilinear
Rectilinear
Rectilinear
Rectilinear
Rectilinear
Bottom-top Rectilinear fill pattern
Rectilinear
Rectilinear
Rectilinear
Rectilinear
Rectilinear
Fill angle
90°
90°
90°
90°
90°
90°
Layer height
0.2mm
0.15mm
0.15mm
0.2mm
0.15mm
0.15mm
Extrusion multiplier
1
1.09
1.09
1.02
1.32
1.1
Nozzle 240–235 °C 205–210 °C 215–210 °C 230–245 °C 245–250 °C 225–230 °C temperature Bed 100–110 °C 60 °C temperature
60 °C
90 °C
75–80 °C
80 °C
Print speed perimeters
45 mm/s
45 mm/s
45 mm/s
20 mm/s
40 mm/s
45 mm/s
Fan speed
15%
10–15%
10–15%
10–15%
10–15%
1.3 Stereolithography Settings (SLA) Similarly, the SLA printing process required adjustments to specific settings for the photopolymer resins. The settings adjusted included layer thickness, curing time, and power intensity (Table 2). 1.4 X-ray Imaging The X-ray imaging experiments were performed at AMIR using a custom spectral photon-counting µCT-scanner. We used a 90 kVp spectrum at ~10 W output power, generated by a Hamamatsu L8106 microfocus tube. The detector employed was an Dectris Eiger2x 1M-W equipped with a 750 µm CdTe sensor. Due to its high stopping power
296
A. Dermitzakis et al. Table 2. Settings of the photopolymers resins for 3D printing technique
Resin
Clear
Purple
Thickness of layer
0.1 mm
0.1 mm
Curing
7 min
10 min
Power intensity
3
3
in combination with a 75 µm pixel pitch, this detector is suitable for mammographic application. The 3D printed cubes were arranged in a row and a motorized stage was used to move the block laterally between acquisitions, ensuring that the same area of the detector was used in each measurement.
Fig. 2. Top: X-ray Imaging setup, Hamamatsu X-ray tube is seen on the right and the Dectris detector on the left. Bottom: Example of radiographic image.
For each printed cube and two open beam positions beyond the ends of the row, ten frames were recorded with a one-second acquisition time per frame, and for each energy threshold (5-79keV, 1keV spacing). Figure 2 shows the experimental setup and an example of the radiographic measurements that were used to derive the attenuation curves. On the right hand side of
‘3D Printed Breast Phantoms Materials for X-ray Imaging Techniques’
297
the radiographic image, the PVC cube supporting the actual sample is seen in black (strongly absorbing). The green box circumscribes the image area that was evaluated. Three large detector gaps can be seen, outside of the measured area (appearing as white vertical bars), while smaller detector gaps are also visible inside in the measured area (thin bright lines). These pixels were excluded from the evaluation.
2 Results The results of the experimental attenuation coefficients for the thermoplastic materials and photopolymer resins are presented (Fig. 3) and discussed in this section. The primary goal is to compare the linear experimental attenuation coefficients with the linear theoretical attenuation coefficients provided by the National Institute of Standards and Technology (NIST)—Physical Measurements Laboratory.
Fig. 3. Attenuation Coefficient (cm-1)—Photon Energy (MeV) graph. Comparison of theoretical attenuation coefficients. The materials are listed by name from top to bottom according to experimental results graph (right): PLA, PLA_Pro, PET_G, Purple, CPE, Clear, ABS, Nylon
Upon examining the total counts graph for all materials, an unexpectedly high variance of count values towards lower energies is observed, which might be attributed to unstable performance of the X-ray source. Therefore, the experimental attenuation coefficients are compared to the theoretical attenuation coefficients in the range of 25 to 79 keV, where the variance appears to obey Poissonian statistics. To determine the attenuation coefficients, counts from neighboring energy thresholds are subtracted to obtain counts per energy interval, and an average open beam value is calculated from open beam measurements taken before and after measurement of the
298
A. Dermitzakis et al.
material samples. The experimental attenuation coefficients of the materials are then calculated using the equation: I = Io e − µx
(1)
From Eq. (1), the attenuation coefficient (µ) can be derived as: µ = 1n [[(Io/I )]]/x
(2)
where: Io represents the average Open, I denotes the subtracted counts, and x is the thickness of the material, which is 2 cm. 2.1 Comparison of the Linear Theoretical and Experimental Attenuation Coefficients The theoretical Attenuation Coefficient was measured according to the National Institute of Standards and Technology (NIST)—Physical Measurements Laboratory for each one of the thermoplastic materials we used for 3D printing. Figure 4 displays a comparison between measured and theoretical spectral attenuation coefficients for the materials investigated in our study. The results indicate a systematic mismatch to the theoretical attenuation profiles, which will need to be investigated further.
3 Conclusion The presented study aimed to investigate the spectral attenuation properties of various thermoplastic materials used for 3D printing by spectral X-ray imaging using photoncounting detectors. However, our preliminary results indicate a strong mismatch between theoretical and measured spectral attenuation profiles, indicating a potential systematic problem with the method employed. The generally observed behavior was that there was a large discrepancy between expected and observed absorption at lower energies, while there was better agreement at higher energies (with the exception of PLA/PLA Pro). The measured absorption at low energies was always considerably below the theoretical values. This might be caused by problems of the experimental design: 1. Charge sharing: In a PCD, the energy of a photon is determined by the amount of charge carriers it generates inside the semiconductor material (here CdTe). When close to a pixel boundary, this charge can be distributed between neighboring pixels, and be falsely detected as multiple low-energy photons instead of one high-energy photon. This artificially increases count rates at low energies and could partly explain the lower-than expected apparent absorption. 2. Compton scattering: For the measured energy range and chemical compositions, the expected absorption is to a large part due to inelastic Compton scattering. The NIST reference data assumes a “thin beam” experiment, i.e. scattered photons are never detected. In our experiment there is a finite solid angle of detection, so photons scattered at small angles, or scattered multiple times inside the sample, can be detected. As the scattering is inelastic, this artificially increases count rates, and thus decreases measured absorption, at lower energies.
‘3D Printed Breast Phantoms Materials for X-ray Imaging Techniques’
299
Fig. 4. Comparison of Experimental and Theoretical Attenuation of all materials.
3. Differences in composition: The study used published generic chemical formulas rather than the (unknown) actual chemical composition used in the production of the filaments, which probably differ to some extent. Additives or chemical modifications for coloring and optimized printing behavior, as well as polymerization catalysts are likely to be present in the filament materials. Some polymers are also hygroscopic and may contain an unknown amount of water. To further clarify these issues, we plan to repeat the current study using a Medipix3 CdTe detector, featuring a so-called Charge Summing Mode, which allows to suppress charge sharing by summing detected signals from adjacent pixels within a certain time window to the pixel with the highest relative contribution.
300
A. Dermitzakis et al.
In conclusion, although the theoretical and experimental attenuation coefficients did not align, the study provides insights into the absorption behavior of thermoplastic materials at various energy levels. The effects of Compton scattering and charge sharing, which could explain the observed deviations from theoretically expected absorption spectra, are also likely to be relevant in actual imaging applications.
References 1. Boone, J.M.: Dedicated breast CT: the past, present, and future of breast imaging. Med. Phys. 44(8), e110–e123 (2017). https://doi.org/10.1002/mp.12277 2. Glick, S.J.: Phantom studies and receiver operating characteristic analysis in nuclear medicine. Semin. Nucl. Med. 44(3), 179–189 (2014). https://doi.org/10.1053/j.semnuclmed. 2013.12.003 3. Ruschin, M., Mainprize, J.G., Mawdsley, G.E., Yaffe, M.J., Jong, R.A.: Design and characterization of a 3D anthropomorphic breast phantom. Med. Phys. 38(7), 4226–4231 (2011). https://doi.org/10.1118/1.3595689 4. Hsu, C.M., Palmeri, M.L., Rosenzweig, S.J.: Fabrication of a tissue-mimicking breast phantom for combined ultrasound, elastography, and MRI. Med. Phys. 40(6), 063501 (2013). https://doi.org/10.1118/1.4803391 5. Sun, Z., Migler, K.B.: 3D printing of interdigitating PLA-PCL structures with tunable mechanical properties. Macromol. Mater. Eng. 298(1), 83–91 (2013). https://doi.org/10.1002/mame. 201200213 6. Solomon, J., Samei, E.: Quantum noise properties of CT images with anatomical textured backgrounds across reconstruction algorithms: FBP and SAFIRE. Medical Physics, 41(9), 091908 (2014). https://doi.org/10.1118/1.4893500. 7. Glick, S.J., Ikejimba, L.C.: Advances in digital and physical anthropomorphic breast phantoms for x-ray imaging. Med. Phy. 45(10), e870-e885 (2018) 8. Chong, A., Weinstein, S.P., McDonald, E.S., Conant, E.F.: Digital breast tomosynthesis: concepts and clinical practice. Radiology 292(1), 1 (2019) 9. Kudelski, R., Cieslik, J., Kulpa, M., Dudek, P., Zagorski, K., & Rumin, R.: Comparison of cost, material and time usage in FDM and SLS 3D printing methods. In 2017 XIIIth international conference on perspective technologies and methods in MEMS design (MEMSTECH) (pp. 12–14). IEEE (2017) 10. Ivanov, D., Bliznakova, K., Buliev, I., Popov, P., Mettivier, G., Russo, P., ... & Bliznakov, Z.: Suitability of low density materials for 3D printing of physical breast phantoms. Phys. Med. Biol. 63(17), 175020 (2018) 11. Malliori, A., Daskalaki, A., Dermitzakis, A., & Pallikarakis, N.: Development of physical breast phantoms for X-ray imaging employing 3D printing techniques. Open Med. Imag. J. 12(1) (2020) 12. NIST Homepage. https://physics.nist.gov/PhysRefData/XrayMassCoef/chap2.html, last accessed 2023/02/20 13. Ballabriga, R., Alozy, J., Campbell, M., Frojdh, E., Heijne, E. H. M., Koenig, T., Llopart, X., Marchal, J., Pennicard, D., Poikela, T., Tlustos, L., Valerio, P., Wong, W., & Zuber, M.: Review of hybrid pixel detector readout ASICs for spectroscopic X-ray imaging. In Journal of Instrumentation (Vol. 11, Issue 1). Institute of Physics Publishing (2016). https://doi.org/ 10.1088/1748-0221/11/01/P01007
Multi-Scale Assessment of Harmonization Efficacy on Resting-State Functional Connectivity Emma Tassi1,2 , Federica Goffi1 , Maria Gloria Rossetti2 , Marcella Bellani3 , Benedetta Vai4,5 , Federico Calesella4,5 , Francesco Benedetti4 , Anna Maria Bianchi1(B) , Paolo Brambilla2,6 , and Eleonora Maggioni1 1 Department of Electronics, Information and Bioengineering, Politecnico Di Milano, Milan,
Italy [email protected] 2 Department of Neurosciences and Mental Health, Fondazione IRCCS Ca’ Granda Ospedale Maggiore Policlinico, Milan, Italy 3 Department of Neurosciences, Biomedicine and Movement Sciences, Section of Psychiatry, University of Verona, Verona, Italy 4 Division of Neuroscience, UnitofPsychiatryandClinicalPsychobiology, IRCCS Ospedale San Raffaele, Milan, Italy 5 Università Vita-Salute San Raffaele, Milano, Italy 6 Department of Pathophysiology and Transplantation, University of Milan, Milan, Italy
Abstract. Raising trends of neuroimaging data sharing among different research centers, including resting-state functional magnetic resonance imaging (rs-fMRI) measurements, have driven to accessible large-scale sample and improvement of reliability and consistency of downstream analyses. However, in this context several concerns arise for non-biological confounding factors mainly related to differences in magnetic resonance scanners and imaging parameters among sites. Until now, there is limited knowledge of the impact of site-to-site variations in rsfMRI functional connectivity (FC) measures and the most suitable harmonization approach for mitigating such impact. In this study, we aimed to quantitatively evaluate the site-to-site variations in rs-fMRI FC patterns and how the widely used ComBat harmonization performs in removing them. A multi-scale analytical approach was adopted, from single pairs of regions to resting-state networks (RSNs) and to the entire brain. Our findings show that ComBat removes unwanted site effects from rs-fMRI FC measures while improving signal-to-noise ratio (SNR) in the data and RSNs identifiability. Further, we identify and visualized specific FC links highly affected by site, highlighting differences in such effects among RSNs. Overall, our findings demonstrate that ComBat is effective in harmonizing rs-fMRI FC measures, emphasizing also the overall RSNs identifiability and the enhancement of the majority of single RSNs in the entire brain connectome. Keywords: Magnetic resonance imaging · Brain · Neuroimaging · Connectome
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 301–308, 2024. https://doi.org/10.1007/978-3-031-49062-0_33
302
E. Tassi et al.
1 Introduction Over the last decades, collecting and sharing large amounts of neuroimaging data are becoming increasingly common and are leading to a growing trend towards forming and collecting data from large-scale, heterogenous samples covering a wide age span. Indeed, integration of structural and functional brain magnetic resonance imaging (MRI) data from multiple sites is useful to increase statistical power related to the sample size and thus would help to obtain larger samples of specific rare disorders. Despite the advantages related to pooling brain MRI data from different centers, multi-site studies are affected by non-biological sources of variability, attributable to the use of different scanners and imaging parameters, affecting reproducibility and consistency of downstream analysis. A recent approach used to deal with site-related heterogeneity is the application of ComBat tool (Combating batch effects) [1], applied as pre-processing stage to remove site’s effects while preserving site-specific biological variability. While effectiveness of ComBat harmonization has been largely explored in multi-site diffusion tensor imaging (DTI) and cortical thickness- and volume-based studies, the knowledges on how site-to-site variations interfere with multi-scale resting-state functional magnetic resonance imaging (fMRI) measures, from single pairs of regions to RSNs and to the entire brain, are still largely limited. Recent multi-studies have demonstrated that ComBat could eliminate differences related to site in fMRI data, enabling a slight reduction of measurement bias and improvement of signal-to-noise ratio (SNR) in the data [2]. In this study, we aimed to quantitatively evaluate the effects of site-specific MRI scanners and sequences on the resting-state brain functional connectivity (FC) patterns at multiple spatial scales, from pairs of regions to RSNs and up to the entire brain, employing resting-state fMRI datasets from three sites. First, we examined the magnitude of site’s effect acting on resting-state FC connections after harmonization, identifying FC links highly affected by site at whole-brain and single RSN-level. Further, we evaluate the novel quality control (QC) metric of functional connectivity con- trast (FCC) [3] in raw and harmonized FC matrices to investigate how data quality changes with respect to the identifiability of RSNs after the application of the ComBat technique. Specifically, FCC metrics were computed at the whole-brain and RSN-levels, in order to evaluate how the ComBat harmonization pipeline is able to highlight each RSN in the overall functional connectome. We hypothesized that (1) substantial site effects are present in the resting-state FC patterns before harmonization, (2) the magnitude of site’s effects changes across different RSNs, (3) ComBat method im- proves the SNR in the data, leading to an increase in the identifiability of large-scale RSNs.
2 Methods 2.1 Subjects and Experimental Protocol The MRI dataset included in our analysis was composed of resting-state fMRI data acquired in three sites from 143 participants, divided in 70 healthy controls (HC) and 73 subjects with a diagnosis of depression. Subjects were aged between 19 and 86 years (mean ± std= 49.04±18.28 years). No statistically significant differences were found in age and sex distributions across sites. The three sites in which the MRI dataset
Multi-Scale Assessment of Harmonization Efficacy …
303
was acquired are the Azienda Ospedaliera Universitaria Integrata of Verona (Site1), Ospedale San Raffaele of Milan (Site2), and Fondazione IRCCS Cà Granda Ospedale Maggiore Policlinico of Milan (Site3). Both T1-weighted MRI and resting-state fMRI data were acquired across the three sites using different 3T MRI scanners, with imaging parameters reported in Table 1. Subjects were instructed to keep their eyes closed, not to think of anything in particular, and not to fall asleep. Table 1. MRI scanner parameters. AOUV: Azienda Ospedaliera Universitaria Integrata of Verona, OSR: Ospedale San Raffaele of Milan and MILANO: Fondazione IRCCS Cà Granda Ospedale Maggiore Policlinico sof Milan. MPRAGE, magnetization-prepared rapid acquisition gradient echo; TFE, Turbo Field Echo; FOV, Field of View; TR, repetition time; TE, time to echo.
Structural
Functional
AOUV (3T Siemens)
OSR (3T Philips)
MILANO (3T Philips Achieva DStream)
3D MPRAGE
3D MPRAGE
3D TFE
Matrix 256 × 256 × 160
Matrix 256 × 256 × 182
Matrix 256 × 256 × 180
FOV 256 mm2
FOV 256 mm2
FOV 256 mm2
Voxel size
Voxel size
Voxel size
1.00 × 1.00.× 1.00
1.00 × 1.00 × 1.00
1.00 × 1.00 × 1.00
TR 2500 ms
TR 2500 ms TE 30 ms
TR 2000 ms
TE 30 ms
TE 30 ms
Flip angle 75°
Flip angle 85°
Flip angle 90°
FOV 210 mm2
FOV 192 mm2
FOV 256 mm2
Matrix 64 × 64 × 36
Matrix 96 × 96 × 38
2.2 MRI Data Pre-Processing The three sites included in the study employed a 3T scanner with different scanner manufactures and sequences applied. The T1-weighted and resting-state fMRI data were pre-processed using the statistical parametric mapping (SPM12) software [4]. For each subject, the pre-processing included spatial realignment of the fMRI volumes, structuralfunctional co-registration, tissue segmentation, spatial normalization to the standard MNI space and spatial smoothing with a 6 mm FWHM Gaussian window. The overall quality of the resting-state fMRI dataset was assessed by monitoring the extent of movement artefact. Specifically, the extent of motion was measured using the framewise displacement (FD) metric [5]. The average FD across included subjects was 0.41 mm (±0.76 mm), below the commonly used FD threshold of 0.5 mm. Moreover, we computed the average FD across subjects of each site (Site 1: 0.49 mm (±0.48 mm), Site 2: 0.15 mm (±0.08 mm), Site 3: 0.37 mm (±0.36 mm)), resulting all below FD threshold.
304
E. Tassi et al.
2.3 Seed-Based Functional Connectivity Analysis Using the SPM Marsbar toolbox [6], we obtained the parcellation of the subjects’ fMRI volumes in N = 90 regions of interest (ROIs) of the Automated Anatomical Labeling (AAL). Cerebellum and vermis AAL ROIs were excluded, since they were not covered by the fMRI volumes in some or all subjects. For each participant, Pearson correlation coefficients between BOLD time series of each pair of ROIs were calculated, resulting in a N × N FC adjacency matrix. From the 90 × 90 FC matrix, we further excluded Gyrus Rectus and Olfactory Cortex ROIs in order to evaluate the FC among areas included at least in one of the seven RSNs considered (Salience network (SAL), Default Mode network (DMN), Central executive network (CEN), Motor network (MOT), Visual network (VIS), Auditory network (AUD) and Basal Ganglia network (BG)), resulting in a 86 × 86 FC matrix. 2.4 ComBat Harmonization of Functional Connectivity Matrix The removal of non-biological site’s effects was performed using the ComBat pipeline for FC matrices [6] adapted to be applied to the upper triangle elements. Connectivity values were harmonized by considering the contribution of age, sex and diagnosis in order to preserve the biological variability in the data. 2.5 Evaluation of Harmonization Effects on Functional Connectivity Values In order to assess any significant site’s effect on group-level FC matrices before and after ComBat harmonization, we applied a one-way ANOVA based on each FC link as dependent variable and as group of independent variables the factors of age, sex, diagnosis and site. The magnitude of site’s effect before and after applying ComBat was extracted based on ANOVA F-score maps associated to the variable of site and evaluated considering the whole FC matrix and the FC extracted at the level of each RSN. Following ComBat, we visualized the FC links both in the whole brain and for each RSN highly associated with site’s effect (i.e., high F-score maps value) and investigated how these effects change after the harmonization. RSNs were also compared considering the percentage of FC links significantly affected by site before and after harmonization. The ANOVA results were corrected using Bonferroni’s correction for multiple comparisons (adjusted p < 0.05). 2.6 Evaluation of Harmonization Effects on Functional Connectivity Quality Control Metric To investigate the harmonization performance in terms of SNR preservation or improvement, as well as of RSNs identifiability, we computed the QC metric of functional connectivity contrast (FCC) before and after harmonization. FCC could be used to evaluate SNR improvement, matched with an improvement in RSNs identifiability, by quantifying the difference in FC values among within-network edges (WNEs) and between-network edges (BNEs). Specifically, an increase in SNR and RSNs identifiability is associated with greater correlation values for WNEs compared to BNEs and thus higher values of
Multi-Scale Assessment of Harmonization Efficacy …
305
FCC. The QC metric of FCC was computed before and after harmonization to evaluate how the pipeline could modify the extent to which WNEs differ from BNEs. In particular, to investigate the FCC increase, associated with an increase of correlation values for WNEs compared to BNEs (i.e., a drop in FC values between networks with respect to within-networks ones), we computed the difference between the FC matrices after versus before harmonization and extracted the FC links associated with a lower value after harmonization (which we hypothesize to be BNEs). FCC was computed both at whole-brain level and at the single RSN-level. Specifically, we extracted WNEs of all the networks and of each network for evaluating the whole-brain RSN and single-RSN identifiability, respectively. As follow, based on Wilcoxon signed-rank tests, we assessed whether the whole-brain and RSN-specific FCC significantly increased or decreased after harmonization (p < 0.05).
3 Results 3.1 ComBat Harmonization Performance on Functional Connectivity The analysis of magnitude of site’s effect before harmonization at the level of each RSN showed that the DMN and CEN were the RSNs with the highest number of FC links significantly affected by site (84% for DMN and 79% for CEN), followed by SAL (54%) and VIS (54%). Figure 1 shows the FC links highly affected by site’s ef- fects by reporting thresholded ANOVA F-score maps (adjusted p 0.05 (continued)
408
Š. Mandal Table 2. (continued)
Free fatty acid
All participants
Diabetics
Controls
P** value
Docosatretraenoic acid C22:4n-6, μmol/L
0.7115 ± 0.00
0.1245 ± 0.00
1.3825 ± 0.00 P < 0.05a
Docosapentenoic acid C22:5n-3, μmol/L
1.7285 ± 0.00
0.0258 ± 0.00
2.7014 ± 0.00 P < 0.05a
Docosahexaenoic acid C22:6n-6, μmol/L
1.0680 ± 0.00
0.3569 ± 0.00
1.8807 ± 0.00 P < 0.01b
* Values are presented as mean ± SD (standard deviation) ** Significance of difference in Student’s t-test (a p < 0.05, b p < 0.01, c p < 0.001)
Fig. 3. Spearman’s correlation coefficient between total bilirubin and myristic acid (rho = − 0.327, p = 0.007; 3A), stearic acid (rho = −0.446, p = 0.000; 3B), and linolenic acid (rho = − 0.227, p = 0.012; 3C), in diabetic patients.
In line with these findings, in recent investigations demonstrated a possible role of individual FFAs and de novo lipogenesis (DNL) in the development of T2D. DNL is a biochemical process that occurs in the liver and involved in the endogenous synthesis of specific FAs, such as palmitic, palmitoleic, stearic and oleic acids and it is connected to the pathophysiology metabolic disorders, including T2D [13–16]. Fumiaki Imamura et al. showed that the concentrations of specific FFAs in the DNL had both, positively and negatively significant association with T2D incidence [17]. Results of previously studies showed that total bilirubin significantly reduced hepatic lipids, triglycerides accumulation, and DNL by increasing the hepatic β-oxidation of free fatty acids and improvement in the liver function. The inverse relationship of serum total bilirubin levels with liver fat accumulation and specific FFAs (SFA and PUFA) has prompted the possibly of bilirubin in treatment for NAFLD [18–21]. Also, increased concentration of total bilirubin may reduce visceral obesity and improve insulin resistance by lowering inflammation in adipose tissue, which make it as a potential therapeutic to protect for high levels of insulin in patients with diabetes or patients with uncontrolled diabetes) [22, 23].
Association Between Serum Free Fatty Acids
409
Fig. 4. Spearman’s correlation coefficient between total bilirubin and myristic acid (rho = − 0.421, p = 0.013; 4A), stearic acid (rho = 0.290, p = 0.035; 4B), arachidic acid (rho = −0.738, p = 0.037; 4C), and Dihomo-γ-linolenic acid (rho = 0.337, p = 0.013; 4D) in control participants. Table 3. Comparison between biochemical parameters in controls subjects and diabetic patients*. Parameter
All participants
Diabetics
Controls
P** value
Z
Glucose, mmol/L
7.85 ± 0.27
9.40 ± 0.34
5.24 ± 0.08
0.000c
4.970
4.79 ± 0.09
0.000c
4.681 1.710
HbA1c, %
6.10 ± 0.12
6.92 ± 0.12
Total Cholesterol, mmol/L
5.29 ± 0.10
5.65 ± 0.14
5.07 ± 0.05
0.006b
HDL-cholesterol, mmol/L
1.34 ± 0.05
1.16 ± 0.06
1.63 ± 0.05
0.000c
3.686
Lauric acid C12:0, 33.910 ± 0.01 μmol/L
62.346 ± 0.02
19.692 ± 0.01
0.010a
1.585
Stearic acid C18:0, 62.628 ± 0.01 μmol/L
83.886 ± 0.01
26.128 ± 0.01
0.003b
1.806
γ-Linolenic acid C18:3n-6, μmol/L
79.861 ± 0.01
176.814 ± 0.04
0.013a
1.581
99.252 ± 0.01
(continued)
410
Š. Mandal Table 3. (continued)
Parameter
All participants
Diabetics
Controls
P** value
Z 2.347
Arachidic acid C20:0, μmol/L
15.765 ± 0.00
7.482 ± 0.00
40.613 ± 0.00
0.000c
Behenic acid C22:0, μmol/L
20.237 ± 0.00
7.0537 ± 0.00
44.567 ± 0.01
0.000c
2.234
Docosapentaenoic acid C22:5n-3, μmol/L
1.284 ± 0.00
0.0858 ± 0.00
2.7014 ± 0.00
0.047a
1.368
Age, years
55 ± 1.20
61 ± 1.18
44 ± 1.61
0.000c
3.452
Total bilirubin, μmol/L
13.01 ± 0.77
14.25 ± 1.24
11.11 ± 0.31
0.002b
1.856
* Values are presented as mean ± SD (standard deviation)
** Significance of difference in Kolmogorov-Smirnov test (Z) (a p < 0.05, b p < 0.01, c p < 0.001)
4 Conclusion In this study it is found a significant association of total bilirubin concentrations and myristic acid, stearic acid, arachidic acid, and dihomoγlinolenic acid in the control group. Serum of total bilirubin levels was also associated with myristic acid, stearic acid, and γlinolenic acid in diabetic patients. It appears that elevated levels of total bilirubin and FFA of different chain lengths and saturation are associated with the risk of development of T2D and could be used as potential T2D biomarkers and therapeutic targets.
References 1. Carracher, A.M., Marathe, P.H., Close, K.L.: International diabetes federation 2017. J. Diabetes 10(5), 353–356 (2018) 2. Mandal, S.: New molecular biomarkers in precise diagnosis and therapy of type 2 diabetes. Heal. Technol. 1–8 (2019). https://doi.org/10.1007/s12553-019-00385-6 3. Dorcely, B., Katz, K., Jagannathan, R., Chiang, S.S., Oluwadare, B., Goldberg, I.J., et al.: Novel biomarkers for prediabetes, diabetes, and associated complications. Diabetes Metab. Syndr. Obes.: Targets Ther. 10, 345–361 (2017) 4. Abbasi, A.: Mendelian randomization studies of biomarkers and type 2 diabetes. Endocr. Connect. 4, 249–260 (2015) 5. Mandal, S., Causevic, A., Dzudzevic Cancar, H., Semiz, S.: Free fatty acid profile in type 2 diabetic subjects with different control of glycemia. In: CMBEBIH International Conference on Medical and Biological Engineering in Bosnia and Herzegovina 2017. IFMBE Proceedings, vol. 62, pp. 781–786. Springer, Heidelberg (2017) 6. Sobczak, A.I.S., Blindauer, A.C., Stewart, J.A.: Changes in plasma free fatty acids associated with type-2 diabetes. Nutrients 11(9), 1–85 (2019)
Association Between Serum Free Fatty Acids
411
7. Inoguci, T., Sonoda N., Maeda Y.: Bilirubin as an important physiological modulator of oxidative stress and inflammation in metabolic syndrome and diabetes: a new aspect on old molecule. Diabetol. Int. 7, 338–341 (2016) 8. Yang, M., Ni, C., Chang, B., Jiang, Z., Zhu, Y., Tang, Y., et al.: Association between serum total bilirubin levels and the risk of type 2 diabetes mellitus. Diabetes Res. Clin. Pract. 152, 23–28 (2019) 9. Žiberna, L., Jenko-Pražnikar, Z., Petelin, A.: Serum bilirubin levels in overweight and obese individuals: the importance of anti-inflammatory and antioxidant responses. Antioxidants 10, 1352 (2021) 10. International Diabetes Federation (IDF): IDF Diabetes Atlas. International Diabetes Federation, Brussels (2019) 11. Jendrassik, L., Grof, P.: Colorimetric method of determination of bilirubin. Biochem. Z. 297, 81–82 (1938) 12. Lepage, G., Roy, C.C.: Specific methylation of plasma nonesterified fatty acids in a one-step reaction. J. Lipid Res. 27, 114–120 (1986) 13. Wei, Y., Liu, C., Lai, F., Dong, S., Chen, H., Chen, L., et al.: Associations between serum total bilirubin, obesity and type 2 diabetes. Diabetol. Metab. Syndr. 143, 1–7 (2021) 14. Hana, C.A., Klebermass, E.-M., Balber, T., Mitterhauser, M., Quint, R., Hirtl, Y., et al.: Inhibition of lipid accumulation in skeletal muscle and liver cells: a protective mechanism of bilirubin against diabetes mellitus type 2. Front. Pharmacol. 11(636533), 1–14 (2021) 15. Zhang, F., Guan, W., Fu, Z., Zhou, L., Guo, W., Ma, Y., et al.: Relationship between serum indirect bilirubin level and insulin sensitivity: results from two independent cohorts of obese patients with impaired glucose regulation and type 2 diabetes mellitus in China. Int. J. Endocrinol. 2020, Article ID 5681296, 10 (2020) 16. Liu, J., Dong, H., Zhang, Y., Cao, M., Song, L., Pan, Q., et al.: Bilirubin increases insulin sensitivity by regulating cholesterol metabolism, adipokines and PPARγ levels. Sci. Rep. 5(09886), 1–11 (2015) 17. Imamura, F., Fretts, A.M., Marklund, M., Ardisson Korat, A.V., Yang, W.-S., Lankinen, M., et al.: Fatty acids in the de novo lipogenesis pathway and incidence of type 2 diabetes: a pooled analysis of prospective cohort studies. PLoS Med. 17(6), 1–17 (2020) 18. Johnston, L.W., et al.: Association of NEFA composition with insulin sensitivity and beta cell function in the Prospective Metabolism and Islet Cell Evaluation (PROMISE) cohort. Diabetologia 61, 821–830 (2018) 19. Li, Q., et al.: Associations between serum free fatty acid levels and incident diabetes in a 3-year cohort study. Diabetes, Metab. Syndr. Obes.: Targets Ther. 14, 2743–2751 (2021) 20. Rebelos, E., Seghieri, M., Natali, A., Balkau, B., Golay, A., Piatti, P.M., et al.: Influence of endogenous NEFA on beta cell function in humans. Diabetologia 58, 2344–2351 (2015) 21. Martins, A.R., Nachbar, R.T., Gorjao, R., Vinolo, M.A., Festuccia, W.T., Lambertucci, R.H., et al.: Mechanisms underlying skeletal muscle insulin resistance induced by fatty acids: importance of the mitochondrial function. Lipids Health Dis. 11(30), 1–11 (2012) 22. Takei, R., Inoue, T., Sonoda, N., Kohjima, M., Okamoto, M., Sakamoto, R., et al.: Bilirubin reduces visceral obesity and insulin resistance by suppression of inflammatory cytokines. PLoS ONE 14(10), e0223302 (2019) 23. Mandal, Š.: The determination of total serum bilirubin concentration in type 2 diabetes patients. Bull. Chemists Technol. Bosnia Herzeg. 56, 7–12 (2020)
HPLC-UV Determination and Comparison of Extracted Corticosteroids Content with Two Methods M. Daci´c1(B) , Alija Uzunovi´c2 , Larisa Alagi´c-Džambi´c3 , and Saša Pilipovi´c2 1 Development and Registration Department, Bosnalijek, Juki´ceva 53, 71000 Sarajevo, Bosnia
and Herzegovina [email protected] 2 Agency for Medicinal Products and Medical Devices, Titova 9, Sarajevo, Bosnia and Herzegovina 3 Quality Assuarance and Quality Control Department, Juki´ceva 53, 71000 Bosnalijek, Bosnia and Herzegovina
Abstract. Counterfeited topical preparations and their presence on the market are the problems in the whole World. There are a great number of manufacturers who are making and selling preparations as natural and pure herbal in which they add corticosteroids. These corticosteroids are not declared and a consumer doesn’t know that these preparations are counterfeited. The aim of this investigation was to examine the influence of different extraction solvents in the analysis of corticosteroids content. The analysis was performed with HPLC-UV Agilent 1220 with manual injection. The first method used reversed-phase Zorbax Phenyl, 250 × 4.6 mm, 5.0 µm column at a wavelength of 240 nm. Injection volume was 20 µL and the oven temperature was 45 °C. The mobile phase was water, HCOOH 0.1%:acetonitrile HCOOH 0.1% with a flow rate fixed at 1.0 mL/min. Extraction was done with pure methanol and a mixture of methanol:water = 50:50. Better extraction with pure methanol was for hydrocortisone; while mometasone was better extracted with a mixture of methanol:water = 50:50 and other corticosteroids had pretty much equal results in both solvents. The second method used a ZORBAX Eclipse XDB-C18, 150 × 4.6 mm, 3.5 µm column at 254 nm. Injection volume was 20 µL and the oven temperature was 25 °C. The mobile phase was acetonitrile:water = 55:45 with a flow rate fixed at 1.0 mL/min. In this work two different methods of extraction were presented and the problems that could appear during the analysis of corticosteroids. Keywords: Corticosteroid · HPLC-UV · Alclometasone dipropionate · Betamethasone dipropionate · Dexamethasone · Hydrocortisone · Mometasone furoate · Clobetasole propionate
1 Introduction Topical corticosteroids are highly effective compounds that are now widely used in dermatology for the treatment of various autoimmune and inflammatory disorders. Many corticosteroids are misused for diverse indications such as pigmentation, acne, pruritus, © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 412–424, 2024. https://doi.org/10.1007/978-3-031-49062-0_45
HPLC-UV Determination and Comparison of Extracted Corticosteroids
413
fungal or bacterial infections, rashes, and numerous other conditions. These products contain various constituents, most of which have hazards or toxic elements [1]. Several steps and improved formulations are constantly evolving in the scientific world, aimed at addressing most of the associated adverse events. However, until the dawn of the ‘ideal topical steroid’, the implications of application misuse will remain significant [2]. Pharmaceutical counterfeiting is becoming a serious problem both in developed and developing countries [3]. Counterfeiting of products is a global problem. As long as clothes, clocks, leatherwear, etc. are fake there is no danger, but when it comes to drugs, counterfeiting can be life-threatening [4]. There are many papers presenting methods and results of different ways of analysis for counterfeited preparations [5–9]. Present research tried to find out the best one under our operative conditions.
2 Materials and Methods In the next passuses, results of two different HPLC methods will be presented which can be used for determination of corticosteroids in ointments. Because of lack of different detectors and large number of columns which can be used for this purpose, the best option that has to be used for determination of corticosteroids were the chosen methods. 2.1 First Method – Operating Conditions 2.1.1 Standard Solutions Stock standard solutions were prepared by weighing 10 mg of standard in a 10-mL volumetric flask and dissolving in methanol. Standards were weighed in separated flasks in order to determinate the retention time for every corticosteroid. After that, dilutions of stock standards for the calibration curve were made. Four standards were mixed, without betamethasone dipropionate, because of overlapping peaks. Mixed standards for calibration curves were prepared by pipetting of 100, 200, 300, 400 and 500 µL in five volumetric flasks of 10 mL and a mixture of methanol:water = 50:50 was added till the mark. Betamethasone standards were prepared separately. 2.1.2 Sample Preparation For the accuracy of the method, six different ointments with the accurately declared content of corticosteroids were used for the determination of corticosteroids. Extractions were done with pure methanol and with a mixture of methanol:water = 50:50. The mass of ointment was weighed in order to obtain the concentration equal to standard 3 (which is used for linearity), once 10 mL of solvent was added and closed with a plug. The mass for every single sample was different because of different concentrations of corticosteroids in the samples. Then, the mixture was heated at 60–70 °C with occasionally mixing, and after that, cooled at the room temperature and centrifuged. The procedure was repeated three times. The extract was filtered through the 0.45 µm nylon filters and injected to the system.
414
M. Daci´c et al.
2.1.3 Method Parameters The method used in this investigation was described in work published in JCS [9]. Instrument: Agilent 1220 Infinity LC, manual injection Column: Zorbax Phenyl, 250 × 4.6 mm, 5.0 µm Flow: 1 mL/min Column temperature: 25 °C Injection volume: 20 µL Wavelength: 254 nm Elution: Time (min)
Water, HCOOH 0.1% (% v/v)
0
95
Acetonitrile, HCOOH 0.1% (%v/v) 5
2
95
5
7
60
40
25
60
40
45
30
70
46
95
5
50
95
5
2.2 Second Method – Operating Procedures 2.2.1 Preparation of Standard Solutions The same stock solutions used for the first method were used for the second method too. Mixed standards used for calibration curves were prepared by pipetting of 100, 200, 300, 400 and 500 µL in five volumetric flasks of 10 mL and added with a mixture of acetonitrile:water:acetic acid = 60%:40%:0.1% to the mark. 2.2.2 Sample Preparation Samples with the exactly declared amount of corticosteroids were purchased from pharmacies and used for checking the accuracy of the method. The preparation procedure was the same as for any kind of herbal sample used for checking the presence of corticosteroids. The sample was weighed to amount of 0.5 g in the test tube with cover in which was added 200 µL of propylparaben–internal standard (IS) and 9.8 mL mixture of acetonitrile:water:acetic acid = 60%:40%:0.1%. The content was heated for 10 min in a water bath at 60 °C, mixed with Vortex for 5 min. The procedure was repeated three times and the content was left in a water bath for 15–20 min. After that, samples were kept in the fridge for 15 min, centrifuged for 10 min at 4000 rpm and filtered through the nylon filter paper (0.45 µm).
HPLC-UV Determination and Comparison of Extracted Corticosteroids
415
2.2.3 Method Parameters The method used in this investigation was combination of three methods in published paper/article below: Rahmayuni at all (2018), Matiˇcevi´c (2017) and Uzunovi´c (2006) [16–18]. Instrument: Agilent 1220 Infinity LC, manual injection Column: ZORBAX Eclipse XDB-C18, 150 × 4.6 mm, 3.5 µm Flow: 1 mL/min Column temperature: 25 °C Injecting volume: 20 µL Wavelength: 254 nm Mobile phase: acetonitrile:MiliQ water = 55:45.
3 Results and Discussion 3.1 Results of the First Method 3.1.1 Chromatographic Separation Peaks are well separated with clearly distinguished retention times for chosen corticosteroids. Betamethasone peak was overlapped with alclomethasone peak, so if both corticosteroids are present in one sample, only qualitative analysis was possible to do. For herbal preparations with not declared presence of corticosteroids, the qualitative analysis should be sufficient for confirmation of counterfeited preparation. Clobetasole propionate was determined just qualitatively because of a lack of clobetasole standard. By comparing the results obtained with this method and results from the other methods, it can be concluded that picks of these corticosteroids have shorter retention time. In these cases, recovery was better too. The average standard deviation for recovery was 0.5% for the extraction with EtOAcs, 0.6% for the extraction with ether, and 0.9% for the extraction with CH2Cl2 in the work of Reepmeyer [10]. Extraction results in two different ointments which contain clobetasole propionate was 96.6% and 94.0% in the work of Yamini, so this corticosteroid has good solubility and is easy to extract [11]. Calibration curves had good linearity and satisfying concentrations range so they could be used for future calculations (Table 1). Table 1. Results from calibration curves for selected corticosteroids (y = ax ± b) Corticosteroid
RT
Coef. corr.
a
b
Betamethasone
38.99
0.9992
38.432
Alclomethasone
38.76
0.9995
37.24
Dexamethasone
13.866
0.9985
48.499
−58.577
Hydrocortisone
12.157
0.9995
52.348
−39.623
Momethasone
40.125
0.9999
54.762
6.746
7.269 13.917
416
M. Daci´c et al.
3.1.2 Determined Content of Corticosteroids in Ointments As discussed before, this was a method for qualitative or quantitative analysis of six different corticosteroids in ointments. In the following, the results of the accuracy and problems that occurred during operation procedures will be presented. Pure methanol gives better extraction results than 50% methanol on one hand, but on the other hand, more problems occurred using this solvent (Table 2). During the extraction and time necessary for cooling it to room temperature, an emulsion which was very difficult to filter. Using a mixture of methanol:water = 50:50, this wasn’t the case, because the water content made it easier to separate the ointment after cooling. The next problem was the retention time which was susceptible to changes, so for this method is very important to have an internal standard - IS (propylparaben) which was not used. Using IS, every change that happened to the sample will reflect to the IS as well. In the book “Steroid analysis” is explained in detail extraction procedures because of the problems which occur with these samples, advising to take an internal standard during analysis from the beginning to the end [12]. A lot of extractions were based on solid-phase extraction which is easier and has better accuracy but these methods are more expensive [13–15]. Table 2. Recovery of the extracted corticosteroids Sample with
%Recovery Solvent: methanol
%Recovery Solvent: 50% methanol
Alclometasone dipropionate
97.60
104.36
Betamethasone dipropionate
100.12
126.29
Dexametasone
100.35
104.20
Momethasone furoate
89.69
97.68
Hydrocortisone
93.34
81.54
3.2 Results of the Second Method 3.2.1 Calibration Curves Peaks are well separated with the different retention times for chosen corticosteroids. Clobetasole propionate was determined just qualitatively because of a lack of clobetasole standard (Table 3). This method has a lot of advantages over the first method, because the time of analysis is shorter (analysis duration is just 12 min), the peak resolution is better, there is no difference between retention times of standards and samples with corticosteroids, and extraction procedure is easier and have a better recovery (Fig. 1). Calibration curves had a good linearity and satisfying concentration range so there could be used for future calculations (formula y = ax ± b) (Table 4).
HPLC-UV Determination and Comparison of Extracted Corticosteroids
417
Table 3. Determined lamda maximum for selected corticosteroids Standard
Lamda max (nm)
Alclomethasone dipropionate (AM)
246.5
Betamethasone dipropionate (BM)
242
Dexamethasone (DM)
241.5
Hydrokortison (HK)
248
Mometasone furoat (MM)
255
Propylparaben (PP)
256
Methylparaben (MP)
251.5
Fig. 1. Chromatogram of mixed corticosteroids – second method
3.2.2 Accuracy Accuracy was done in two ways (Tables 5, 6, 7, 8, 9, 10, 11, 12, 13 and 14): 1. About 1 g of comfrey fat (Gavez) was weighed three times in three separate test tubes with a 25 mL cap. 100, 200, and 300 µL of each corticosteroid standard and 200 µL of the propylparaben (IS) standard were pipetted into the given test tubes, and 9.3, 8.8, and 8.3 mL of the basic solvent were added, respectively, and closed with a
418
M. Daci´c et al. Table 4. Results from calibration curves for selected corticosteroids
Corticosteroid
RT
Coef. corr.
a
b
Betamethasone
9.451
0.999
27.312
−19.209
Alclomethasone
7.734
0.9994
31.28
−24.216
Dexamethasone
2.017
0.9996
33.374
−2.157
Hydrocortisone
1.746
0.9998
39.985
7.734
Momethasone
8.741
0.9998
56.394
−18.485
stopper. The samples were treated according to the “Sample preparation” procedure. Samples are recorded three times. 2. Fats with a precisely defined amount of corticosteroids were taken, and a mass of fat whose content corresponds to the concentration of standard number 3 (100%) was weighed into marked test tubes with stoppers. In addition to this, the same number of other test tubes were taken in which the appropriate mass of fat with declared corticosteroids was weighed in the same way, as well as a similar mass of the comfrey fat sample, in order to determine possible interference from placebo or other substances present. 10 mL of the basic solvent was added, and the test tubes were closed with a cap. Samples are recorded three times.
Table 5. Accuracy and precision uncertain (Up) for alclometasone dipropionate (spiked samples) Alclometasone
AVG
STDEV
RSD%
Up%
% through cal. curve
105.61
1.00
0.94
1.97
% through RF
103.74
1.08
1.04
1.29
97.74
0.28
0.29
1.07
% through ISTD
Table 6. Determination of alclometasone dipropionate in market samples calculated by three different ways Alclometasone
% through RF
% through cal. curve
% through ISTD
Market sample (MS)
97.27
94.79
–
MS + Gavez
93.00
91.44
–
MS + ISTD
94.75
92.25
92.57
3.2.3 LOD LOQ From the stock solution 100 µL of each standard as well as 200 µL of internal standard is pipetted into a 10 mL measuring vessel. The same is supplemented with the basic solvent
HPLC-UV Determination and Comparison of Extracted Corticosteroids
419
Table 7. Accuracy and precision uncertain (Up) for betamethasone dipropionate (spiked samples) Betamethasone
AVG
STDEV
RSD%
Up%
% through cal. curve
98.79
1.06
1.09
1.88
% through RF
96.72
1.13
1.23
2.26
% through ISTD
94.31
1.19
1.34
1.93
Table 8. Determination of betamethasone dipropionate in market samples calculated by three different ways Betamethasone Market sample (MS)
% through RF
% through cal. curve
% through ISTD
99.70
98.26
–
MS + Gavez
110.54
107.76
–
MS + ISTD
101.79
100.29
100.41
Table 9. Accuracy and precision uncertain (Up) for dexametasone (spiked samples) Dexametasone
AVG
STDEV
RSD%
Up%
% through cal. curve
103.74
0.81
0.78
0.80
% through RF
104.14
0.82
0.78
0.72
99.41
0.09
0.09
0.99
% through ISTD
Table 10. Determination of dexametasone in market samples calculated by three different ways Dexametasone
% through RF
% through cal.curve
Market sample (MS)
93.62
93.04
% through ISTD –
MS + Gavez
91.90
91.38
–
MS + ISTD
95.30
94.83
90.55
Table 11. Accuracy and precision uncertain (Up) for hydrocortisone (spiked samples) Hydrocortisone
AVG
STDEV
RSD%
Up%
% through cal. curve
101.41
2.04
1.99
0.81
% through RF
101.62
2.00
1.92
1.05
% through ISTD
102.13
1.28
1.20
1.32
420
M. Daci´c et al.
Table 12. Determination of hydrocortisone in market samples calculated by three different ways Hydrocortisone
% through RF
% through cal. curve
% through ISTD
Market sample (MS)
101.97
103.14
–
MS + Gavez
97.87
98.65
–
MS + ISTD
94.99
95.77
95.55
Table 13. Accuracy and precision uncertain (Up) for momethasone furoate (spiked samples) Momethasone
AVG
STDEV
RSD%
Up%
% through cal. curve
105.78
2.35
2.17
1.13
% through RF
105.72
2.45
2.26
0.67
% through ISTD
106.12
3.86
3.52
0.99
Table 14. Determination of momethasone furoate in market samples calculated by three different ways Momethasone
% through RF
% through cal.curve
% through ISTD
Market sample (MS)
93.98
92.73
–
MS + Gavez
94.23
93.24
–
MS + ISTD
94.10
93.35
95.62
up to the mark. 0.8; 0.6; 0.4; 0.2 and 0.1 mL is pipetted from the dilution solution in the other six measuring vessels of 10 mL, and they are topped up with the basic solvent up to the mark (Table 15). Table 15. Determined LOD and LOQ Corticosteroid
LOD (µg/mL)
LOQ (µg/mL)
AM
0.6788
2.0363
BM
0.8042
2.4126
DM
0.1443
0.4329
HK
0.3152
0.9456
MM
0.2756
0.8267
Clobetasole
Nd
Nd
HPLC-UV Determination and Comparison of Extracted Corticosteroids
421
3.2.4 Precision - Repeatability In 10 mL measuring vessels, pipette 300 µL of each stock standard, including IS, and top up the same with the basic solvent up to the mark. The prepared solution is filtered and recorded seven times for repeatability and five times for robustness, under different conditions (Table 16). Table 16. Precision - repeatability Corticosteroid
Up% through RF
Up% through cal. curve
Up% through ISTD
AM
0.3169
0.3008
0.0238
BM
0.3169
0.3044
0.0306
DM
0.3215
0.3202
0.0137
HK
0.3193
0.3232
0.0369
MM
0.3090
0.3018
0.0703
Clobetasole
–
–
–
3.2.5 Precision - Reproducibility The precision in terms of reproducibility was performed using propylparaben, considering that it is a method in which the changes on propylparaben are the same as those on the active substance. 21 recordings of the solution were performed under reproducibility conditions, by pipetting 0.2 mL of the basic standard and diluting it with the basic solvent to 10 mL. In the same way, a propylparaben solution with a higher concentration was recorded 12 times more, which was prepared by pipetting 0.3 mL of the standard and diluting it to 10 mL. RSD% (21 recordings) = 1.41% for samples with lower propylparaben concentrations RSD% (7 recordings) = 0.87% for samples with higher propylparaben concentrations. 3.2.6 Robustness A standard sample was prepared in the concentration value of standard number 3 with the presence of an internal standard, and it was recorded under five different conditions (Table 17).
422
M. Daci´c et al. Table 17. Corticosteroids RSD% calculated at three different ways
Corticosteroid
RSD% through RF
RSD% through cal. curve
RSD% through ISTD
AM
3.9006
3.7011
0.4713
BM
3.9051
3.7490
0.4333
DM
3.9530
3.9381
0.5357
HK
3.9774
4.0264
0.5586
MM
3.8607
3.7710
0.2844
Clobetasole
–
–
–
Terms: R1 - Share of mobile phases: water:acetonitrile - 56:44 R2 - Share of mobile phases: water:acetonitrile - 54:46 R3 - mobile phase flow - 0.95 ml/min R4 - mobile phase flow - 1.05 ml/min R5 - Column temperature - 30 °C. 3.2.7 Determined Content of Corticosteroids in Ointments Recovery of analyzed samples was satisfying and had range 93–101% which is in correlation with regulatory requirements. In Table 18. Were showed the results of samples with a declared amount of corticosteroids. Table 18. Recovery of the extracted corticosteroids Sample with
%Recovery
Alclometasone dipropionate
97.3657
Betamethason dipropionate
99.7985
Dexametasone
93.6237
Momethasone furoate Hydrocortisone
93.9794 101.9724
With this method, 24 herbal samples with not declared corticosteroids, intended for the treatment of skin diseases, were analyzed. Eight had a presence of corticosteroids which concentration was showed in Table 19.
HPLC-UV Determination and Comparison of Extracted Corticosteroids
423
Table 19. Results of analysed herbal samples Sample
mg/g
Corticosteroid found
U6
0.0171 ± 0.0021
Momethasone furoate
U11
0.1609 ± 0.0288
Betamethasone dipropionate
U12
0.0086 ± 0.0007
Hydrocortsone
U13
3.4068 ± 0.3475
Dexametasone
U17
0.2360 ± 0.0241
Dexametasone
U22
0.1905 ± 0.0340
Betamethasone dipropionate
U23
0.0110 ± 0.0020
Betamethasone dipropionate
U24
0.0039 ± 0.0003
Dexametasone
4 Conclusion With the first method, it is possible to determinate a very large number of corticosteroids, but if it’s intend to analyse a few corticosteroids, it is not recommended. Some multiple problems occur such as extraction of corticosteroids, filtration of samples, and the changing of retention times of corticosteroids. This method is good for qualitatively analysis and results showed pretty good extraction, but the time of analysis isn’t practical (it lasts 50 min, which is too long). The second method is much easier, extraction is very good, the analysis is not long (12 min) and there is a possibility of analyzing a very big number of corticosteroids – it is analyzed a six of them plus propylparaben, but there is a possibility to increase this number of corticosteroids. Peaks are separated with stable retention times and recovery is in range 93–101% which is excellent. The method is selective and sensitive so with this method you can determinate a very small amount of corticosteroids which is great for finding counterfeited topical preparations.
References 1. Jaccob, A.A., Yaqoub, A.A., Rahmani, M.A.: Impact of abuse of topical corticosteroids and counterfeit cosmetic products for the face: prospective demographic study in Basrah City, Iraq. NCBI 15(1), 25–31 (2020). https://doi.org/10.2174/1574886314666191001100357. https:// www.ncbi.nlm.nih.gov/pubmed/31573892 2. Nnoruka, E.N., Daramola, O.O.M., Ike, S.O.: Misuse and abuse of topical steroids: implications. Expert Rev. Dermatol. 2(1), 31–40 (2007). https://doi.org/10.1586/17469872.2. 1.31. https://www.tandfonline.com/doi/abs/10.1586/17469872.2.1.31?src=recsysIn&journa lCode=ierg20& 3. Deisingh, A.K.: Pharmaceutical counterfeiting, RSC. Analyst (3) (2005). https://doi.org/10. 1039/B407759H. https://pubs.rsc.org/en/content/articlelanding/2005/an/b407759h/unauth#! divAbstract 4. Holzgrabe, U., Malet-Martino, M.: Analytical challenges in drug counterfeiting and falsification—the NMR approach. J. Pharm. Biomed. Anal. 55(4), 679–687 (2010). https://www. sciencedirect.com/science/article/abs/pii/S0731708510007223
424
M. Daci´c et al.
5. Fiori, J., Andrisano, V.: LC–MS method for the simultaneous determination of six glucocorticoids in pharmaceutical formulations and counterfeit cosmetic products. J. Pharm. Biomed. Anal. 91, 185–192 (2014). https://www.sciencedirect.com/science/article/abs/pii/S07317085 13006110 6. McEwen, Elmsjö, A., Lehnström, A., Hakkarainen, B., Johansson, M.: Screening of counterfeit corticosteroid in creams and ointments by NMR spectroscopy. J. Pharm. Biomed. Anal. 70, 245–250 (2012). https://www.sciencedirect.com/science/article/abs/pii/S07317085 12003809 7. Giaccone, V., Polizzotto, G., Macaluso, A., Cammilleri, G., Ferrantelli, V.: Determination of ten corticosteroids in illegal cosmetic products by a simple, rapid, and high-performance LC-MS/MS method., IJAA 2017, Article ID 3531649, 12 (2017). https://doi.org/10.1155/ 2017/3531649 8. Jalili, R., Miraghaeia, S., Mohamadib, B., Babaeia, A., Bahramia, G.: Detection of corticosteroid compounds and phosphodiesterase inhibitors (PDH-5) as counterfeit in herbal products available in Iranian market by HPLC method. JRPS 4(1), 75–81 (2015). http://citeseerx.ist. psu.edu/viewdoc/download?doi=10.1.1.854.8045&rep=rep1&type=pdf 9. Gimeno, P., et al.: HPLC–UV method for the identification and screening of hydroquinone, ethers of hydroquinone and corticosteroids possibly used as skin-whitening agents in illicit cosmetic products. J. Chromatogr. Sci. Chromatogr. Sci. 2015, 1 (2015). https://doi.org/10. 1093/chromsci/bmv147 10. Reepmeyer, J.C.: Screening for corticosteroids in topical pharmaceuticals by HPLC with a scanning ultraviolet detector. J. Liquid Chromatogr. Related Technol. (2001). https://doi.org/ 10.1081/JLC-100103404 11. Yamini, Y., Safari, M., Shamsayei, M.: Simultaneous determination of steroid drugs in the ointment via magnetic solid phase extraction followed by HPLC-UV. J. Pharm. Anal. (2018). https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6104151/ 12. Makin, H.L.J., Gower, D.B., Kirk, D.N.: Steroid Analysis—Book. Springer-Science + Business Media, University of London (1995) 13. Bonazzi, D., Andrisano, V., Gatti, R., Cavrini, V.: Analysis of pharmaceutical creams: a useful approach based on solid-phase extraction (SPE) and UV spectrophotometry. J. Pharm. Biomed. Anal. 13, 1321–1329 (1995). https://doi.org/10.1016/0731-7085(95)01536-T 14. Merck, Determination of Hydrocortisone from Topical Cream Using Discovery DSC-Si SPE and Reversed-Phase HPLC-UV. https://www.sigmaaldrich.com/technical-documents/art icles/analytical/solid-phase-extraction/determination-hyrdocortisone-topical-cream.html 15. Johnston, S.E., Gill, L.N., Wei, Y.-C., Markovich, R., Rustum, A.M.: Development and Validation of a Stability-Indicating RP-HPLC Method for Simultaneous Assay of Betamethasone Dipropionate, Chlorocresol, and for the Estimation of Betamethasone Dipropionate Related Compounds in a Pharmaceutical Cream and Ointment (2010) 16. Rahmayuni, E., Harmita, H., Suryadi, H.: Development and validation method for simultaneous analysis of retinoic acid, hydroquinone and corticosteroid in cream formula by highperformance liquid chromatography. J. Appl. Pharm. Sci. 8(09), 087–092 (2018). https://doi. org/10.7324/JAPS.2018.8913. http://www.japsonline.com 17. Matiˇcevi´c, M.: Utjecaj ekstrakcijskog otapala na odredivanje odabranih kortikosteroida u prirodnim dermatološkim proizvodima – diplomski rad, Farmaceutsko-biokemijski fakultet, Zagreb (2017) 18. Uzunovi´c, A.: Specijalistiˇcki rad, Zavod za kontrolu lijekova FBiH, Sarajevo (2006)
Can Microneedles Revolutionize Ocular Drug Delivery? Merima Šahinovi´c1(B) , Amina Tucak-Smaji´c1 , Kenan Muhamedagi´c2 , Lamija Hindija1 , Ognjenka Rahi´c1 , Jasmina Hadžiabdi´c1 , Edina Vrani´c1 , ˇ c2 and Ahmet Ceki´ 1 Faculty of Pharmacy, Department of Pharmaceutical Technology, University of Sarajevo,
Zmaja od Bosne 8, 71000 Sarajevo, Bosnia and Herzegovina [email protected] 2 Faculty of Mechanical Engineering, Department of Machinery Production Engineering, University of Sarajevo, Vilsonovo šetalište 9, 71000 Sarajevo, Bosnia and Herzegovina [email protected]
Abstract. The anterior and posterior segments of the eye are extremely difficult to reach and treat, due to their barrier properties and inaccessibility of the ocular tissues. Given that traditional drug delivery systems, such as topical eye drops, or hypodermic needles, possess many disadvantages such as causing pain, and discomfort or resulting in low drug bioavailability, microneedles (MNs) are recently proposed as minimally invasive drug delivery systems for treating ocular diseases. These systems enable the localization of drugs within the ocular tissue, while on the other hand being painless. Given that MNs will definitely change how eye medication formulations are delivered in the future, this work aimed to summarize the previous studies on the potential of using microneedles as drug delivery systems for treating ocular diseases and outline the future aspect of MNs production technologies. Keywords: Drug delivery systems · Eye diseases/drug therapy · Microneedles
1 Introduction Blindness and visual impairment may be the most serious health issue in the world. According to the World Health Organization (WHO), there are 285 million visually impaired persons in the world of which 39 million are blind and 246 have limited vision [1]. In general, ocular diseases are classified as anterior (glaucoma, uveitis, dry eye syndrome, bacterial or fungal keratitis) and posterior segment diseases (diabetic macular edema, diabetic retinopathy, or other chorioretinal diseases) [2]. Due to the delicate nature of ocular tissues, it is challenging to deliver drugs using traditional methods such as topical application, systemic administration, and periocular and intravitreal routes. Given that conventional hypodermic needles that are used as intraocular injection causes pain and discomfort, as well as the potential for ocular complications, microneedles have emerged as a minimally-invasive drug delivery system for ocular delivery [2, 3]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 425–433, 2024. https://doi.org/10.1007/978-3-031-49062-0_46
426
M. Šahinovi´c et al.
2 Methods Comprehensive data search has been conducted in the relevant databases based on the selected keywords (“microneedles in ocular drug delivery” OR “microneedles for ocular drug delivery” OR “microneedle-based ocular drug delivery systems” OR “ocular drug delivery using microneedles”). Only original scientific papers published in English were included. Review articles and articles related to other administration routes were excluded. We also excluded papers with incomplete in vivo/in vitro characterization, or lack of parameters relevant for dataset building.
3 Microneedles Microneedles (MNs) represent a third generation of minimally invasive devices that were initially developed to improve the transdermal delivery of various drugs. During the administration, they only minimally interact with the dermal pain receptors (nociceptors) before puncturing the stratum corneum layer of the epidermis [4, 5]. MNs are composed of micron-sized (25–2000 µm) solid or hollow projections and can be fabricated from a variety of materials, including silicon, stainless steel, ceramic, glass, sugar, metal, and polymers [2]. As MNs successfully delivered drugs into both posterior and anterior segments of the eye, they have recently emerged as delivery systems for targeted transscleral delivery of drugs into ocular tissues [6]. Due to the optimal length of MNs, the epithelial transport barriers and conjunctival clearance mechanism can be bypassed using MNs, which minimizes the potential of retinal damage and enable targeted-drug deposition within ocular tissues [2, 7]. In addition, their design enables hygienic, safe, and user-friendly applications [8]. Three types of MNs are usually used for ocular drug delivery. The first type, coated MNs, penetrates ocular tissue followed by the dissolving of drugs in a few minutes after application, after which the device is removed [7]. Hollow MNs are used for drug delivery from the external reservoir into ocular tissue by passive diffusion or application of pressure [9–12]. Finally, dissolving MNs are composed of a soluble matrix, which entirely dissolves after administration [13]. With these systems, it is possible to deliver various agents including microparticles and nanoparticles, depot-forming gels, or drug solutions [2].
4 Microneedle Systems as Ocular Drug Delivery Systems Recent research has demonstrated that MNs can be used to treat a variety of ocular illnesses, including glaucoma, uveitis, retinal vascular occlusion, age-related macular degeneration retinitis pigmentosa, etc. The scientific interest in ocular MNs systems is steadily expanding as seen by the abundance of published research papers. Overall, it appears that MNs will completely change how eye medication formulations are delivered in the future [14]. However, all of these systems are still in the early development stage, and there are no available commercial products on the market.
Can Microneedles Revolutionize Ocular Drug Delivery?
427
In 2007, Jiang at el. Demonstrated for the first time that coated MN can serve as an effective ocular delivery system via intrascleral and intracorneal routes. Results showed that MNs successfully penetrated the human cadaveric sclera, and 30 s after penetration, the drug coating on the needles quickly disintegrated within the scleral tissue. The concentration of the drug in the anterior chamber was found to be 60 times higher after in vivo delivery via fluorescein-coated microneedles. The most significant observation was the absence of any discernible inflammatory reactions following the implantation of MNs [15]. Song et al. designed a stainless steel solid MNs pen system, a single MN with a spring-loaded MN applicator to provide impact insertion. This system was made using a transfer molding process and was shown to successfully deliver rhodamine dye deep enough to reach the stromal layer of the cornea. This was an indication that more targeted, less invasive drug delivery may be accomplished with MN pan [16]. Hollow MNs were also shown as a minimally invasive manner for rapid drug delivery to the eye. It was found that individual MN can infuse 10–35 µl of the drug solution into the sclera. The solution can contain microparticles as well as nanoparticles thus presenting the opportunity for controlled-release drug delivery [17]. Patel et al. revealed in their research for the first time that hollow MNs can serve as a platform to deliver nanoparticle and microparticle suspensions into the suprachoroidal space of rabbit, pig, and human eyes. Using a minimum MN length of 800 µm and a minimum pressure of 250 kPa, 20 and 100 nm particles were successfully injected. This method of injecting particles may provide sustained drug distribution from the suprachoroidal region and thereby decrease dosage frequency [18]. The same research group utilized hollow MNs to target the suprachoroidal space of the rabbit eye in vivo and found that concentrations of injected materials were 10-fold or greater in the ocular posterior segment compared with the anterior segment could be achieved [19]. Based on a study by Gilger et al. delivery of triamcinolone acetonide into the suprachoroidal space of porcine eyes using hollow MNs was efficient, well-tolerated, and had a good safety profile for up to 3 days [20]. Dissolving polymeric MNs, made of biodegradable, biocompatible polymers, were suggested for ocular drug delivery to address the issues associated with coated and hollow MNs in terms of their manufacturing, precision, or safety concerns. Dissolving microneedles (MNs) made of biodegradable PVP with different MW were used to enhance ocular drug delivery of macromolecules. Dextran, fluorescein sodium, and fluorescein isothiocyanate were used as model macromolecules. Studies conducted in vitro revealed that using MNs significantly improved macromolecule penetration through both the corneal and scleral tissues as compared to topically administered aqueous solutions [21]. Penetration of antibiotics into and through the cornea was improved by rapidly dissolving polymeric MNs array. Using polymeric MNs besifloxacin was effectively delivered through the cornea and results showed that the application of MNs for 5 min significantly (p < 0.05) improved the besifloxacin deposition and permeation through the cornea compared with free besifloxacin solution [22].
428
M. Šahinovi´c et al.
Than et al. manufactured corneal patches with self-implantable needle-shaped micro reservoirs. The biphasic release kinetics and packing of numerous drugs for synergistic treatment are made possible by the double-layer structured MNs (Fig. 1.d). They demonstrate the higher efficacy of such an eye patch in the treatment of corneal neovascularization in comparison to topical eye drops of anti-angiogenic monoclonal antibodies [23]. MN ocular patches containing PVA and PVP polymeric MNs with pilocarpine were successfully developed by Roy et al. When this ocular patch (Fig. 1.a) was applied instead of pilocarpine solution, considerably more pilocarpine permeated across the excised cornea in terms of both cumulative amount and flow, thus proving its effectiveness [24]. A dissolving MN patch based on Poly(D,L-lactide) (PLA) and hyaluronic acid (HA) was successfully manufactured by Shi et al. to deliver fluconazole. MNs patch was able to penetrate the corneal epithelial layer without any obvious ocular irritation, and offer higher drug bioavailability. Superior therapeutic outcome was found in the rabbit model compared with the conventional eye drop formulation [25]. Albadr et al. showed that polymeric rapidly dissolving MNs can serve as a delivery system for amphotericin B in the therapy of fungal infections (Fig. 1.e). After successful penetration into porcine corneal tissue, amphotericin permeation was enhanced. Interestingly, PVP and HA decreased amphotericin cytotoxicity and preserved antifungal activity when compared to the free drug [26]. Datta et al. effectively designed a polymeric PVP MN ocular patch to deliver the high molecular weight molecule cyclosporine A to the cornea. After penetration of MNs into excised porcine cornea, cyclosporine A was released in a controlled manner. The flux and retention of model drug in the porcine cornea were significantly enhanced when compared with commercially available eye drops [27]. Lately, Roy et al. reported manufacturing polymeric MN patches for scleral application to deliver drugs to the posterior region of the eye. They used triamcinolone acetonide (TA) as the model drug and fabricated a microneedle scleral patch (MSP) (Fig. 1.c) as well microneedle corneal patch (MCP) (Fig. 1.b). After penetration into corneal and scleral tissue MNs dissolved within the 60s. The results revealed that MSP may be employed as a minimally invasive drug delivery method to transport drugs to the posterior segment of the eye, as MSP demonstrated considerably higher TA disposition in excised porcine eye globe compared with MCP and TA nanosuspension eye drops [28]. Amer et al. fabricated a photo-responsive hydrogel MN array that, upon swelling, may self-adhere to the application site and that, upon illumination with light, can deswell for simple removal. Interestingly, molds for MNs were fabricated using digital light processing-based 3D printing [29]. The types, materials, dimensions, and shapes of MNs currently in research as ocular drug delivery devices are summarized in Table 1. All of the abovementioned research that mainly discusses formulation development and in vivo outcomes using animal models is encouraging and paves the way for further research to identify the most significant system constraints and additional formulation development. Due to the ocular tissues’ incredibly fragile structure, relative inaccessibility, and barrier characteristics, drug administration to the eye is extremely difficult.
Can Microneedles Revolutionize Ocular Drug Delivery?
429
Fig. 1. a. Stereomicroscopic image of the MNs patch. Reproduced with permission from [24], Taylor & Francis, 2020 b. photograph of MCP and c. photograph of MSP. Reproduced with permission from [28], Elsevier, 2022 d. illustration of eye-contact patch-self-implantable microdrug-reservoirs. Reproduced with permission from [23] Nature Portfolio, 2018 e. light microscopy images of polymeric MN arrays. Reproduced with permission from [26], Springer Nature, 2022.
Traditional ocular drug delivery systems face different drawbacks such as low or subtherapeutic drug levels in the case of topical eyedrops or the need for surgical implantation in the case of implants. On the other hand, using conventional hypodermic needles
430
M. Šahinovi´c et al.
Table 1. Summary of MNs types, materials, dimensions, and shapes used in ocular drug delivery Microneedle type/material
Dimensions/shape
Main result
References
Solid (metal) coated
500–750 µm
Mechanically strong enough to penetrate the human cadaveric sclera; the drug in the anterior chamber was found to be 60x higher after in vivo delivery in rabbits
[15]
Hollow (glass)
200–300 µm elliptical tip opening
Were able to insert into the [17] sclera; nanoparticle suspensions and microparticles were delivered into the sclera
Hollow (glass)
800–1000 µm
One MN injection can disperse a liquid formulation circumferentially around the eye and into the suprachoroidal area
Dissolvable polymeric (polyvinylpyrrolidone, PVP)
800 × 300 µm, conical shaped
MNs penetrated ocular [21] tissue and could rapidly dissolve to form a depot within the tissue and showed a higher degree of permeation of macromolecules
Dissolvable polymeric (polyvinyl alcohol and PVP)
1500 × 400 µm
MN dissolved rapidly [22] within the cornea and the deposition of besifloxacin was significantly greater; the MN device resulted in superior infection control in both in vitro and ex vivo
[18]
(continued)
for intraocular delivery is highly invasive, and can cause discomfort, and pain, thus lowering patient compliance. Thus, MNs as minimally-invasive means of ocular drug delivery can offer a solution to all the abovementioned problems. Less tissue trauma, less drug dosage, and precise localization of the medication make this therapeutic modality suitable for clinical practice. We believe that innovating MNs production technologies such as 3D printing will
Can Microneedles Revolutionize Ocular Drug Delivery?
431
Table 1. (continued) Microneedle type/material
Dimensions/shape
Main result
References
Dissolvable polymeric
521 ± 10 µm in length pyramidal-shaped
The flux of pilocarpine across excised cornea was significantly greater after the application of MNs compared with solution formulation; a study in porcine eye globe: significantly greater availability in aqueous humor
[24]
Dissolvable polymeric (PVP)
545 ± 8 µm length and 279 ± 26 µm width at the base
Greater disposition of TA [28] in the anterior segment of the eye after MCP application; greater TA disposition in the posterior segment of the eye after MSP application on the sclera
Dissolvable polymeric (PVP and hyaluronic acid)
750 × 300 µm, conical shaped
Lower toxicity, faster [26] dissolution, full penetration through corneal tissues with 100% drug recovery, and, retained antifungal activity
Dissolvable polymeric (PVP K-90/K-30, PVA and PEG-400)
535 ± 15.7 and 287 ± 2.00 µm conical shaped
Enhanced the retention of drugs in excised porcine cornea; porcine eye globe distribution studies showed a disposition of 0.2% of CsA within the cornea, aqueous humor, lens, vitreous humor, sclera, and choroid-retinal complex
[27]
largely simplify MNs production [30], and together with artificial intelligence offer enormous potential for the development of smart ocular systems that would allow patients to self-administer drugs without any discomfort and pain.
432
M. Šahinovi´c et al.
5 Conclusion MNs are an attractive technology that might provide a minimally invasive way to administer drugs to both anterior and posterior segments of the eye, causing less tissue damage and less pain, and requiring less drug dosage. MNs are advantageous for targeted drug delivery because they permit precise injections within the delicate ocular tissues (such as the cornea and sclera). Finally, MNs have the potential to revolutionize ocular drug delivery, however, for the technology to be transferred from laboratory settings to clinical conditions, the researchers still have a lot of work to do. Acknowledgment. This research has been supported by Federal Ministry of Education and Science, Bosnia and Herzegovina (project number 05-35-1928-1/21).
References 1. He, Y., Nie, A., Pei, J., Ji, Z., Jia, J., Liu, H., et al.: Prevalence and causes of visual impairment in population more than 50 years old: the Shaanxi eye study. Medicine (Baltimore) 99, e20109 (2020) 2. Thakur Singh, R.R., Tekko, I., McAvoy, K., McMillan, H., Jones, D., Donnelly, R.F.: Minimally invasive microneedles for ocular drug delivery. Expert Opin. Drug Deliv. 14, 525–537 (2017) 3. Prausnitz, M.R., Jiang, J., Patel, S.R., Gill, H.S., Ghate, D., McCarey, B.E., et al.: Ocular drug delivery using microneedles. Invest. Ophthalmol. Vis. Sci. 48, 3191 (2007) 4. Jamaledin, R., You, C.K.Y., Zare, E.N., Niu, L.N., Vecchione, R., Chen, G., et al.: Advances in antimicrobial microneedle patches for combating infections. Adv. Mater. 32, 1–29 (2020) 5. Tucak, A., Sirbubalo, M., Hindija, L., Rahi´c, O., Hadžiabdi´c, J., Muhamedagi´c, K., et al.: Microneedles: characteristics, materials, production methods and commercial development. Micromachines 11, 961 (2020) 6. Thakur, R.R.S., Fallows, S.J., McMillan, H.L., Donnelly, R.F., Jones, D.S.: Microneedlemediated intrascleral delivery of in situ forming thermoresponsive implants for sustained ocular drug delivery. J. Pharm. Pharmacol. 66, 584–595 (2014) 7. Khandan, O., Kahook, M.Y., Rao, M.P.: Fenestrated microneedles for ocular drug delivery. Sens. Actuators B Chem. 223, 15–23 (2016) 8. Goyal, G., Garg, T., Rath, G., Goyal, A.K.: Current nanotechnological strategies for treating glaucoma. Crit. Rev. Ther. Drug Carrier Syst. 31, 365–405 (2014) 9. Davis, S.P., Martanto, W., Allen, M.G., Prausnitz, M.R.: Hollow metal microneedles for insulin delivery to diabetic rats. IEEE Trans. Biomed. Eng. 52, 909–915 (2005) 10. Martanto, W., Moore, J.S., Couse, T., Prausnitz, M.R.: Mechanism of fluid infusion during microneedle insertion and retraction. J. Control. Release 112, 357–361 (2006) 11. Wang, P.M., Cornwell, M., Prausnitz, M.R.: Minimally invasive extraction of dermal interstitial fluid for glucose monitoring using microneedles. Diabetes Technol. Ther. 7, 131–141 (2005) 12. Bal, S.M., Ding, Z., Van Riet, E., Jiskoot, W., Bouwstra, J.A.: Advances in transcutaneous vaccine delivery: do all ways lead to Rome? J. Control. Release 148, 266–282 (2010) 13. Larrañeta, E., Lutton, R.E.M., Woolfson, A.D., Donnelly, R.F.: Microneedle arrays as transdermal and intradermal drug delivery systems: materials science, manufacture and commercial development. Mater. Sci. Eng. R Rep. 104, 1–32 (2016)
Can Microneedles Revolutionize Ocular Drug Delivery?
433
14. Rahi´c, O., Tucak, A., Omerovi´c, N., Sirbubalo, M., Hindija, L., Hadžiabdi´c, J., et al.: Novel drug delivery systems fighting glaucoma: formulation obstacles and solutions. Pharmaceutics 13, 1–58 (2021) 15. Jiang, J., Gill, H.S., Ghate, D., McCarey, B.E., Patel, S.R., Edelhauser, H.F., et al.: Coated microneedles for drug delivery to the eye. Investig. Ophthalmol. Vis. Sci. 48, 4038–4043 (2007) 16. Song, H.B., Kim, J.H., Ju, K., Seo, I.H., Lee, J.Y., Lee, S.M., et al.: Impact insertion of transfer-molded microneedle for localized and minimally invasive ocular drug delivery. J. Control. Release 209, 272–279 (2015) 17. Jiang, J., Moore, J.S., Edelhauser, H.F., Prausnitz, M.R.: Intrascleral drug delivery to the eye using hollow microneedles. Pharm. Res. 26, 395–403 (2009) 18. Patel, S.R., Lin, A.S.P., Edelhauser, H.F., Prausnitz, M.R.: Suprachoroidal drug delivery to the back of the eye using hollow microneedles. Pharm. Res. 28, 166–176 (2011) 19. Patel, S.R., Berezovsky, D.E., McCarey, B.E., Zarnitsyn, V., Edelhauser, H.F., Prausnitz, M.R.: Targeted administration into the suprachoroidal space using a microneedle for drug delivery to the posterior segment of the eye. Invest. Ophthalmol. Vis. Sci. 53, 4433–4441 (2012) 20. Gilger, B.C., Abarca, E.M., Salmon, J.H., Patel, S.: Treatment of acute posterior uveitis in a porcine model by injection of triamcinolone acetonide into the suprachoroidal space using microneedles. Investig. Ophthalmol. Vis. Sci. 54, 2483–2492 (2013) 21. Thakur, R.R.S., Tekko, I.A., Al-Shammari, F., Ali, A.A., McCarthy, H., Donnelly, R.F.: Rapidly dissolving polymeric microneedles for minimally invasive intraocular drug delivery. Drug Deliv. Transl. Res. 6, 800–815 (2016) 22. Bhatnagar, S., Saju, A., Cheerla, K.D., Gade, S.K., Garg, P., Venuganti, V.V.K.: Corneal delivery of besifloxacin using rapidly dissolving polymeric microneedles. Drug Deliv. Transl. Res. 8, 473–483 (2018) 23. Than, A., Liu, C., Chang, H., Duong, P.K., Cheung, C.M.G., Xu, C., et al.: Self-implantable double-layered micro-drug-reservoirs for efficient and controlled ocular drug delivery. Nat. Commun. 9, 4433 (2018) 24. Roy, G., Galigama, R.D., Thorat, V.S., Garg, P., Venuganti, V.V.K.: Microneedle ocular patch: fabrication, characterization, and ex-vivo evaluation using pilocarpine as model drug. Drug Dev. Ind. Pharm. 46, 1114–1122 (2020) 25. Shi, H., Zhou, J., Wang, Y., Zhu, Y., Lin, D., Lei, L., et al.: A rapid corneal healing microneedle for efficient ocular drug delivery. Small 18, 2104657 (2022) 26. Albadr, A.A., Tekko, I.A., Vora, L.K., Ali, A.A., Laverty, G., Donnelly, R.F., et al.: Rapidly dissolving microneedle patch of amphotericin B for intracorneal fungal infections. Drug Deliv. Transl. Res. 12, 931–943 (2022) 27. Datta, D., Roy, G., Garg, P., Venuganti, V.V.K.: Ocular delivery of cyclosporine A using dissolvable microneedle contact lens. J. Drug Deliv. Sci. Technol. 70, 103211 (2022) 28. Roy, G., Garg, P., Venuganti, V.V.K.: Microneedle scleral patch for minimally invasive delivery of triamcinolone to the posterior segment of eye. Int. J. Pharm. 612, 121305 (2022) 29. Amer, M., Ni, X., Xian, M., Chen, R.K.: Photo-responsive hydrogel microneedles with interlocking control for easy extraction in sustained ocular drug delivery. J. Eng. Sci. Med. Diagn. Ther. 5, 011001 (2022) 30. Sirbubalo, M., Tucak, A., Muhamedagic, K., Hindija, L., Rahi´c, O., Hadžiabdi´c, J., et al.: 3D Printing—a “touch-button” approach to manufacture microneedles for transdermal drug delivery. Pharmaceutics 13, 924 (2021)
Analysis of Lysozyme as Biomarker in Saliva Emina Haskovi´c1(B) , Lejda Uzunovi´c1 , Tanja Duji´c1 , Aziz Šukalo2 , Meliha Mehi´c2 , Maja Malenica1 , Tamer Bego1 , Neven Meseldži´c1 , Selma Imamovi´c Kadri´c1 , and Una Glamoˇclija1,2 1 Faculty of Pharmacy, University of Sarajevo, Zmaja od Bosne 8, 71 000 Sarajevo, Bosnia and
Herzegovina [email protected] 2 Bosnalijek JSC, Juki´ceva 53, 71 000 Sarajevo, Bosnia and Herzegovina
Abstract. Non-invasive biomarkers are of crucial importance in diagnostics. Their discoveries bring novel opportunities for easy screening and monitoring of diseases. Saliva is one of the abundant sources of non-invasive biomarkers. Although there are efforts to promote development of techniques for using salivary components as biomarkers, still there are many drawbacks that need to be overcome. Potential biomarker in saliva is lysozyme, one of the most important antimicrobial proteins in human body. In the literature, there are several protocols for measurement of lysozyme concentration and activity in saliva. However, little is known about differences in results depending on the saliva collection and handling. The aim of this study was to evaluate how saliva preparation procedures affect results of lysozyme activity measurement when using turbidimetric method. Unstimulated saliva was collected from 20 individuals. Lysozyme activity was analyzed in fresh and saliva samples after freezing at −20 °C for up to ten days. Each sample was analyzed as whole saliva or supernatant after centrifugation. Viscosity and turbidity of saliva were different between fresh samples and samples used after freezing and thawing process. There was statistically significant two-way interaction between freezing and centrifugation of samples on lysozyme activity (p = 0.004). Although the higher lysozyme activity was found in the frozen compared to fresh samples, it did not reach statistical significance. No difference in lysozyme activity was seen between whole and centrifuged samples. Results of this study should be considered when analyzing lysozyme in saliva or interpreting results of various studies that measured salivary lysozyme activity. Keywords: Lysozyme · Saliva · Turbidimetric method
1 Introduction Saliva is a valuable resource for clinical and diagnostic applications. With advantages of non-invasive collection and availability, this type of sample is used in analysis of various conditions [1]. In diagnostics, usually unstimulated saliva is used since the concentration of biomarkers can be decreased in stimulated saliva [2]. One of the abundant antibacterial components in saliva is lysozyme. It participates in the construction of anatomical © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 434–442, 2024. https://doi.org/10.1007/978-3-031-49062-0_47
Analysis of Lysozyme as Biomarker in Saliva
435
and physiological barriers and plays important role in nonspecific defense mechanisms against microorganisms. Lysozyme participates in maintaining the balance in the oral microbiome [3]. The concentration of lysozyme in saliva is influenced by various factors. In healthy individuals, lysozyme activity depends on metabolome and microbiome characteristics [4]. Also, stress [5, 6] or viewing of a humorous videotape [7] were previously shown to affect concentration of lysozyme in human saliva. Reported average lysozyme concentrations in healthy individuals vary between studies. Hankiewicz and Swierczek reported 8.80 (± standard deviation (SD) 2.97) µg/ml in unstimulated saliva [8], Rudney 3.94 (± SD 5.35) µg/ml in stimulated parotid saliva [9], Yeh et al. 1.34 (± SD 0.09) µg/ml in stimulated parotid and 5.76 (± SD 0.33) µg/ml in stimulated submandibular/sublingual saliva [10]. Lysozyme could be used as a biomarker of various conditions such as sarcoidosis [11], intrathoracic lymphadenopathy [12], leukemias [13, 14], and rheumatoid arthritis [15]. Concentration of salivary lysozyme is decreased in periodontitis [16], psoriasis [17], diabetes [18] and increased in COVID-19 [19], Sjögren’s syndrome [20], coronary heart disease [21], and hypertension [22]. Therefore, lysozyme activities could be affected by inclusion and non-inclusion criteria for each study. Another important factor that can influence lysozyme activity is procedure for saliva collection, processing and analysis [23, 24]. So far, there are only several studies evaluating effects of sample handling and used methodology on activity of lysozyme in human saliva. Jenzano et al. found differences between lysozyme activity of same samples analyzed by lysoplate and turbidimetric methods. Also, they found higher lysozyme activity in whole compared to centrifuged saliva [23]. Virella and Goudswaard used lysoplate and immunonephelometric methods to evaluate concentration of lysozyme in saliva samples handled with four different procedures. While filtration of saliva through 0.45 µm filter resulted in complete loss of lysozyme activity, centrifuged saliva had lower lysozyme activity compared to the whole saliva. The highest activity of lysozyme was observed in acidified and dialyzed saliva (protocol used for removal of mucins) with significantly higher lysozyme concentration (19.6 ± 5.4 µg/ml) compared to untreated fresh saliva (13.8 ± 3.7 µg/ml). Parotid stimulated saliva was not suitable sample due to instability of lysozyme [25]. In the literature, no studies that compare lysozyme activity in fresh and frozen saliva samples were found. Although Ng et al. evaluated stability of lysozyme in saliva stored at − 30 °C from one to 12 months, they did not compare the results in frozen versus fresh samples [26]. The aim of this study was to evaluate differences in salivary lysozyme activity in samples from same individuals taken at the same timepoint analyzed as whole or centrifuged saliva and as fresh samples or after freezing at −20 °C for up to ten days.
2 Materials and Methods 2.1 Sample Collection and Storage At least two hours before sampling, participants did not consume food or drink. Unstimulated saliva samples were collected by spontaneous salivation into sterile 1.5 ml tubes. Two samples were collected from each participant. Immediately after collection samples
436
E. Haskovi´c et al.
were placed on ice. One sample was used as a whole saliva while the other was centrifuged for 10 min at 1.500 × g and 4 °C with Mikro 22R centrifuge (Hettich, Germany). Each sample was analyzed fresh or after storage for up to ten days at −20 °C. 2.2 Chemicals Microccocus luteus ATCC No 4698 (Sigma-Aldrich, USA) suspension in phosphate buffer solution was used as a substrate for turbidimetric method. It was freshly prepared for each experiment. Phosphate buffer solution (pH 6.2) was prepared weekly by mixing monosodium phosphate and disodium phosphate solutions in distilled water and adjusting pH to 6.2 by addition of 0.1 M sodium hydroxide or 0.1 M phosphoric acid. All chemicals for phosphate buffer solution were purchased from Sigma Aldrich, Germany. Hen egg white lysozyme (HEWL) standard was purchased from Lysozyme Manufacturing Company B.V. (LMC) with declared 42571 FIP units/mg solid. HEWL dry standard was resuspended in phosphate buffer solution in eight concentrations (starting from 4000.00 FIP units/ml with serial dilutions 1:1 until 31.25 FIP units/ml). For each experiment, fresh solutions of standards were prepared and used. Hydrochloric acid was purchased from Sigma Aldrich, Germany. 2.3 Turbidimetric Method for Lysozyme Activity Analysis The modified turbidimetric method from Japanese Pharmacopoeia 17th edition [27] was used for the analysis of lysozyme activity. Absorbance of substrate was adjusted to 0.800 ± 0.050 at 640 nm. After that, 160 µl of substrate was added to each well of 96-well plate except wells for blank samples where 160 µl of phosphate buffer solution was added. Plate was incubated 5 min at 35 °C. After that, 40 µl of saliva samples or standard solutions of HEWL were added into adequate wells. 40 µl of phosphate buffer solution was added into negative control and blank wells. Plate was shaken for 5 s and incubated 10 min at 35 °C. After that, 20 µl of 1M HCl was added into each well, plate was shaken for 5 s and absorbance was read at 640 nm at spectrophotometer ELx800 (BioTek, USA). The results for saliva samples were presented in FIP units/ml as calculated from HEWL standard curve. Phosphate buffer solution was used as a negative control. 2.4 Ethics Statement The study was approved by the Ethics committee of University of Sarajevo-Faculty of Pharmacy, number 0101-6987/22. The Helsinki Declaration from 1975 and its amendments from 1983 were followed in all procedures. 2.5 Statistical Analysis Linear regression was used for calculation of HEWL standard curve parameters. Normality of data distribution was evaluated with the Shapiro-Wilk’s test. The Kruskal-Wallis H test was used to compare lysozyme activity in four types of samples in preliminary analyses. Two-way repeated measures ANOVA was used for comparison of freezing
Analysis of Lysozyme as Biomarker in Saliva
437
and centrifugation effects on lysozyme activity in saliva samples. Bonferroni correction was applied for multiple testing. All tests were two-sided with p < 0.05 accepted as a statistically significant difference if no Bonferroni correction was applied. Statistical analysis was performed using the SPSS (Statistical Package for Social Sciences) program version 23.0 and using R Statistical Software (Foundation for Statistical Computing, Vienna, Austria) version 4.2.2. Visualization was performed by using ggplot2 [28] and viridis packages.
3 Results Samples from 20 participants were collected. Five samples were very dense and could not be pipetted or centrifuged so they were not analyzed. Range from 0 to 1000 FIP units/ml of HEWL standard gave linear results with coefficient of correlation R2 = 0.9891. For each independent experiment, curve was prepared by using serial dilutions of HEWL standard. Preliminary analysis was performed on two samples. No statistically significant difference was found between four sample types (whether they were centrifuged or not and whether they were fresh or frozen) (p > 0.05 for all performed analyses) (Fig. 1).
Fig. 1. Lysozyme activity in two saliva samples used for preliminary analysis.
After preliminary analysis, 15 samples were analyzed for lysozyme activities after four procedures of preparation (Fig. 2). Viscosity and turbidity of saliva were different between fresh samples and samples used after freezing and thawing process, as assessed visually. In some samples opalescence was observed in supernatant after centrifugation. A two-way repeated measures ANOVA was used for determination of freezing and centrifugation effects on lysozyme activity in saliva. Lysozyme activity was normally distributed as assessed by Shapiro-Wilk’s test of normality on the studentized residuals (p > 0.05). No outliers were identified after examination of studentized residuals for values greater than ±3. There was a statistically significant two-way interaction between freezing and centrifugation of samples on lysozyme activity F(1,14) = 11.842, p =
438
E. Haskovi´c et al.
Fig. 2. Lysozyme activity in 15 samples prepared by four different procedures (fresh or frozen at −20 °C for up to ten days, supernatant after centrifugation or without centrifugation).
0.004. Simple main effects of sample freezing and centrifugation on lysozyme activity were evaluated. In this analysis, Bonferroni correction was applied and p < 0.006 was considered statistically significant. Although the higher lysozyme activity was found in the frozen compared to fresh samples (Table 1), it did not reach statistical significance after Bonferroni correction. Centrifuged saliva after freezing compared to the fresh one had a mean difference of 75.53 FIP units/ml (95% confidence interval (CI), 17.37–133.69) with F(1,14) = 7.76, p = 0.015 (Table 1). Out of 15 centrifuged samples, ten had increased lysozyme activity for more than 20 FIP units/ml after freezing (Fig. 2). Whole saliva after freezing compared to the fresh one had a mean difference of 109.89 FIP units/ml (95% CI, 19.59–200.19) with F(1,14) = 6.813, p = 0.021 (Table 1). Out of 15 whole saliva samples, ten had increased lysozyme activity for more than 20 FIP units/ml after freezing (Fig. 2). No difference in lysozyme activity was seen between whole and centrifuged samples. Frozen centrifuged saliva compared to the frozen whole saliva samples had F(1,14) = 2.282, p = 0.153 while fresh whole saliva compared to the fresh centrifuged saliva samples had F(1,14) = 0.241, p = 0.631 (Table 1).
Analysis of Lysozyme as Biomarker in Saliva
439
Table 1. Lysozyme activity in 15 samples prepared by four different procedures (fresh or frozen at −20 °C for up to ten days, without centrifugation or supernatant after centrifugation). Results are presented as mean value ± standard deviation (SD). Sample type
Lysozyme activity (FIP units/ml)
Centrifuged saliva, fresh
219.99 ± 150.98
Centrifuged saliva, frozen
295.52 ± 159.66
Whole saliva, fresh
232.11 ± 112.31
Whole saliva, frozen
342.00 ± 182.51
4 Discussion Sample preparation influenced the results of lysozyme activity in unstimulated whole saliva. There was statistically significant two-way interaction between freezing and centrifugation of samples on lysozyme activity (p = 0.004). Although no statistical significance was reached, there was a trend of higher lysozyme activity in frozen compared to fresh samples. Since our samples were frozen at −20 °C for up to ten days, we could not expect that lysozyme activity will change. Ng et al. found that lysozyme concentration remains stable from one to three months when centrifuged saliva is stored at −30 °C. However, they did not compare activity of lysozyme in frozen versus fresh samples [26]. We found higher lysozyme activity in ten out of 15 samples after freezing and thawing procedure. This could be due to the presence of aggregates in fresh samples in both whole and centrifuged saliva. Despite centrifugation, supernatant of some samples was opalescent. Noble used repeated centrifugation to obtain clear supernatant for turbidimetric analysis of fresh saliva [29]. In our study, after freezing and thawing procedure, opalescence was lost and there was a precipitate at the bottom or top, depending on the sample. This was the case only with part of samples while others had clear supernatant upon visual inspection after centrifugation. Similarly, Eltze et al. found that saliva viscosity is dramatically decreased after freezing and thawing procedure. Viscosity of saliva depends on various factors such as acute exercise and acrylic dentures. High viscosity of saliva should be considered when developing methods for salivary biomarkers [30]. Opalescence of our samples could interfere with turbidimetric method that is based on the action of lysozyme against microorganism Microccocus luteus. Destruction of microorganisms’ membranes results in absorbance decrease as measured at 640 nm [31]. In samples that were not clear in fresh state but became clear after freezing and thawing procedure, higher decrease of absorbance can be anticipated. To our knowledge, there are no studies comparing concentration or activity of lysozyme in frozen versus fresh samples from same individuals taken at the same timepoint. Procedures for saliva collection and analysis are different in literature. Hankiewicz and Swierczek used the agarose gel diffusion method to evaluate the concentration of lysozyme in centrifuged unstimulated saliva stored at 6 °C for up to one day and found that in healthy persons concentration of lysozyme in saliva and serum was similar. Average concentration of lysozyme in 35 samples was 8.80 (± SD 2.97) µg/ml [8]. Rudney used modified lysoplate method and found average concentration of lysozyme
440
E. Haskovi´c et al.
3.94 (± SD 5.35) µg/ml in 198 samples of stimulated parotid saliva after freezing at − 70 °C [9]. However, Ng et al. found average lysozyme concentration 249.8 µg/ml (from 59.7 to 1062.3 µg/ml) after storage of centrifuged unstimulated saliva for one month at −30 °C [26]. We found no difference in lysozyme activity between whole and centrifuged samples, in both fresh and saliva after freezing and thawing process. On the other side, Jenzano et al. [23] and Virella and Goudswaard [25] found higher lysozyme activity in whole compared to centrifuged saliva. Centrifugation protocol could have influenced the results. Jenzano et al. stored whole saliva at −20 °C (duration of storage was not mentioned in the article). After thawing they centrifuged saliva at 5.000 × g for 4 min at 4 °C [23]. Virella and Goudswaard centrifuged fresh whole saliva at 14.000 rpm for 30 min at 4 °C [25]. Freezing before centrifugation or longer centrifugation period could have affected the results of lysozyme activity. Also, centrifugation speed could have influenced the results. While we centrifuged our samples 10 min at 1.500 × g, both Jenzano et al. and Virella and Goudswaard used higher centrifugation speeds [23, 25]. We found wide variation of lysozyme activity between individuals, even with the same sample preparation procedure. Also other authors found high variability in lysozyme activity between healthy individuals [23, 29, 32]. Limitations of this study were small sample size and lysozyme activity determination by only one method based on turbidity of samples. Five out of 20 samples in our study were not analyzed since they could not be centrifuged due to density. This could be overcome by additional protocol optimization involving dilution for very dense and viscous samples. We found variability in viscosity and turbidity of saliva samples between individuals but also between fresh and frozen samples, as assessed visually. More precise measurement of viscosity and turbidity of saliva samples should be included in further research. Also, different centrifugation protocols should be evaluated and their influence on results analyzed.
5 Conclusion Activity of lysozyme in saliva is influenced by protocol of sample processing. Centrifugation and freezing of samples had two-way interaction with lysozyme activity. When results from different studies evaluating lysozyme concentrations or activities in saliva are compared, the protocol for collection and preparation of samples should be taken into consideration.
References 1. Nunes, L.A.S., Mussavira, S., Bindhu, O.S.: Clinical and diagnostic utility of saliva as a non-invasive diagnostic fluid: a systematic review. Biochem. Med. (Zagreb) 25, 177–192 (2015) 2. Streckfus, C.F., Bigler, L.R.: Saliva as a diagnostic fluid. Oral Dis. 8, 69–76 (2002) 3. Sun, H., et al.: Salivary secretory immunoglobulin (SIgA) and lysozyme in malignant tumor patients. Biomed. Res. Int. 2016, e8701423 (2016) 4. Zaura, E., Brandt, B.W., Prodan, A., et al.: On the ecosystemic network of saliva in healthy young adults. ISME J. 11, 1218–1231 (2017)
Analysis of Lysozyme as Biomarker in Saliva
441
5. Perera, S., Uddin, M., Hayes, J.A.: Salivary lysozyme: a noninvasive marker for the study of the effects of stress on natural immunity. Int. J. Behav. Med. 4, 170–178 (1997) 6. Yang, Y., et al.: Self perceived work related stress and the relation with salivary IgA and lysozyme among emergency department nurses. Occup. Environ. Med. 59, 836–841 (2002) 7. Perera, S., Sabin, E., Nelson, P., Lowe, D.: Increases in salivary lysozyme and IgA concentrations and secretory rates independent of salivary flow rates following viewing of a humorous videotape. Int. J. Behav. Med. 5, 118–128 (1998) 8. Hankiewicz, J., Swierczek, E.: Lysozyme in human body fluids. Clin. Chim. Acta 57, 205–209 (1974) 9. Rudney, J.D.: Relationships between human parotid saliva lysozyme lactoferrin, salivary peroxidase and secretory immunoglobulin A in a large sample population. Arch. Oral Biol. 34, 499–506 (1989) 10. Yeh, C.-K., Dodds, M.W.J., Zuo, P., Johnson, D.A.: A population-based study of salivary lysozyme concentrations and candidal counts. Arch. Oral Biol. 42, 25–31 (1997) 11. Tomita, H., et al.: Serum lysozyme levels and clinical features of sarcoidosis. Lung 177, 161–167 (1999) 12. Lodha, S.C., Mir, M.A.: Usefulness of serum lysozyme measurement in diagnosis of intrathoracic lymphadenopathy. Tubercle 61, 81–85 (1980) 13. Das, S., Basu, A., Mitra, A., Chatterjee, A., Mitra, S.: A study on the lysozyme pattern in murine leukaemia and lymphomas. Curr. Sci. 49, 423–425 (1980) 14. Saito, N.: Correlation between Intracellular and extracellular lysozyme in acute (myelo)monocytic leukemia. Leuk. Lymphoma 2, 347–350 (1990) 15. Pruzanski, W., Saito, S., Ogryzlo, M.A.: The significance of lysozyme (muramidase) in rheumatoid arthritis. I. Levels in serum and synovial fluid. Arthritis Rheum. 13, 389–399 (1970) 16. Ito, T., et al.: Relationship between antimicrobial protein levels in whole saliva and periodontitis. J. Periodontol. 79, 316–322 (2008) 17. Gasior-Chrzan, B., Falk, E.S.: Lysozyme and IgA concentrations in serum and saliva from psoriatic patients. Acta Derm. Venereol. 72, 138–140 (1992) 18. Olayanju, O.A., et al.: Oral innate immunity in patients with type 2 diabetes mellitus in a tertiary hospital in Ibadan Nigeria: a cross-sectional study. Pan Afr. Med. J. 43, 134 (2022) 19. Qaysar Musa, S., Mohammed AliJassim, M., Mohammed Mahmood, M.: Tracing some salivary immune elements in Iraqi SARS-2 patients. Arch. Razi Inst. 77, 1587–1591 (2022) 20. Ryu, O.H., Atkinson, J.C., Hoehn, G.T., Illei, G.G., Hart, T.C.: Identification of parotid salivary biomarkers in Sjogren’s syndrome by surface-enhanced laser desorption/ionization time-offlight mass spectrometry and two-dimensional difference gel electrophoresis. Rheumatology (Oxford) 45, 1077–1086 (2006) 21. Janket, S.-J., et al.: Salivary lysozyme and prevalent coronary heart disease: possible effects of oral health on endothelial dysfunction. Arterioscler. Thromb. Vasc. Biol. 26, 433–434 (2006) 22. Qvarnstrom, M., et al.: Salivary lysozyme and prevalent hypertension. J. Dent. Res. 87, 480– 484 (2008) 23. Jenzano, J.W., Hogan, S.L., Lundblad, R.L.: Factors influencing measurement of human salivary lysozyme in lysoplate and turbidimetric assays. J. Clin. Microbiol. 24, 963–967 (1986) 24. Brandtzaeg, P., Mann, W.V.: A comparative study of the lysozyme activity of human gingival pocket fluid, serum, and saliva. Acta Odontol. Scand. 22, 441–455 (1964) 25. Virella, G., Goudswaard, J.: Measurement of salivary lysozyme. J. Dent. Res. 57, 326–328 (1978) 26. Ng, V., Koh, D., Fu, Q., Chia, S.-E.: Effects of storage time on stability of salivary immunoglobulin A and lysozyme. Clin. Chim. Acta 338, 131–134 (2003)
442
E. Haskovi´c et al.
27. Japanese Pharmacopoeia 17th Edition | Pharmaceuticals and Medical Devices Agency. https:// www.pmda.go.jp/english/rs-sb-std/standards-development/jp/0019.html. Accessed 28 Mar 2023 28. Wickham, H.: Ggplot2: Elegant Graphics for Data Analysis. Springer, New York (2016) 29. Noble, R.E.: Salivary alpha-amylase and lysozyme levels: a non-invasive technique for measuring parotid vs submandibular/sublingual gland activity. J. Oral Sci. 42, 83–86 (2000) 30. Eltze, L., Eltze, M., Garcia, A., Eltze, L., Eltze, M., Garcia, A.: Variability of saliva viscosity— potential impact. Oral health care—an important issue of the modern society (2020). https:// doi.org/10.5772/intechopen.93933 31. Mörsky, P.: Turbidimetric determination of lysozyme with Micrococcus lysodeikticus cells: reexamination of reaction conditions. Anal. Biochem. 128, 77–85 (1983) 32. Prodan, A., et al.: A study of the variation in the salivary peptide profiles of young healthy adults acquired using MALDI-TOF MS. PLoS ONE 11, e0156707 (2016)
Characterization of the Chemical Substance Niacinamide M. Emira1 , M. Daci´c1,2(B) , and Alija Uzunovi´c3 1 Faculty of Pharm and Health, Slavka Gavranˇci´ca 17C, Travnik, Bosnia and Herzegovina
[email protected]
2 Development and Registration Department, Bosnalijek, Juki´ceva 53, 71000 Sarajevo, Bosnia
and Herzegovina 3 Agency for Medicinal Products and Medical Devices, Titova 9, Sarajevo, Bosnia and
Herzegovina
Abstract. Niacinamide is a form of vitamin B3 and belongs to the group of B complex vitamins. It is used as a supplement and medicine, but also for dermocosmetic purposes. The aim of this work was to characterize the chemical substance niacinamide with a declared content of 99.0–101.0%, for the purpose of confirming the identification and the declared degree of purity. The substance was identified according to Ph. Yug. IV and according to Ph. Eur. 9, and the content according to Ph. Eur. 9, whereby the substance satisfied the necessary requirements. The identification was also performed by UV-VIS spectroscopy, where the absorption spectrum of the chemical substance was confirmed compared to the standard one, at 262 nm, and further tests were performed at wavelengths close to the absorption maximum. For the purpose of characterization, a series of 5 solutions of standard substance and 10 solutions of a chemical substance, concentration 0.04 mg/ml, was prepared and their absorption was recorded. For each wavelength at which the recording of the chemical substance was performed, a good recovery with RSD ≤ 2% was confirmed, and the average content of the chemical substance was 101.0%, which corresponds to the declared content. The test also confirmed the linearity of the UV-VIS method in the concentration range 0.005–0.05 mg/ml, with an acceptable correlation coefficient R2 . The test confirmed the identity and content of the chemical in relation to the certified standard substance, and a satisfactory linearity method was developed that can be used in further tests of monocomponent preparations with niacinamide. Keywords: Niacinamide · Characterization · UV-VIS spectroscopy · Absorbance
1 Introduction Niacniamide is a form of vitamin B3 and belongs to the group od vitamin B complex. It is found in different types of food, but it is also used as a food supplement, as a medicine and in cosmetology [1]. Vitamin B3 deficiency causes a systemic disease known as pellagra, which causes symptoms such as dermatitis, diarrhea and dementia, with the possibility of death, depending on the severity and duration of the disease [2]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 443–451, 2024. https://doi.org/10.1007/978-3-031-49062-0_48
444
M. Emira et al.
Niacinamide is an integral part of certain coenzymes, such as niacinamide adenine dinucleotide (NAD+) and its reduced form (NADH), and niacinamide adenine dinucleotide phosphate (NADP+) and its reduced form (NADPH) [3]. Niacinamide in the human body has the same activity as nicotinic acid, but has different pharmacological effects and side effects [2]. The use of niacinamide as a dietary supplement has numerous beneficial effects on the body, and its role is indispensable in preserving the integrity of the skin. The use of niacinamide prevents lipid peroxidation and protects the body from reactive oxygen species, due to its pronounced antioxidant properties [4]. Due to its anti-inflammatory properties, it can be used in various inflammatory conditions of the skin [2], and in doses of 500– 1000 mg/day it reduces the risk of developing skin cancer, except melanoma [5]. It acts as a chemo- and radio-sensitizing agent, improving blood flow in the tumor, thereby reducing tumor hypoxia. This agent also inhibits poly(ADP-ribose) polymerases, enzymes involved in the rejoining of DNA strand breaks induced by radiation or chemotherapy [6]. Niacinamide (Fig. 1) is a white crystalline substance, odorless and tasteless [7]. It dissolves well in water, glycerol, moderately in ethanol, while it is almost insoluble in ether [8]. Numerous methods have been developed for the identification and determination of niacinamide in various matrices, including UV-VIS spectrophotometry, volumetric methods, HPLC, electrochemical methods, etc. [9].
Fig. 1. Chemical structure of niacinamide
The aim of this work was the identification of a chemical substance according to Ph. Yug. IV, and identification and determination according to Ph. Eur. 9. Also, the characterization of the chemical substance was carried out in relation to the standard substance, by confirming the identity and composition of the chemical substance in relation to the declared data, based on various sources and characterization methods used in practice. Also, the linearity of the UV-VIS spectrophotometric method for the determination of niacinamide in monocomponent preparations was tested.
Characterization of the Chemical Substance Niacinamide
445
2 Materials and Methods 2.1 Chemicals Niacinamide USP reference standard and the chemical substance niacinamide (purity 99.0–101.0%) manufactured by Sigma Aldrich, USA, were used for the purpose of the tests. Ethanol, ether, NaOH, CuSO4 , acetic acid, acetic anhydride, perchloric acid and crystal violet were also used for the analysis. 2.2 Identification According to Ph. Yug. IV The look of the substance, its solubility in water, ethanol and ether were examined. Also, identification was made according to Ph. Yug. IV: the substance was heated (identification A), 20 mg was mixed with 5 ml of NaOH (r.o.) and heated to boiling (identification B), 0.1 g was dissolved in 2 ml of water and 2 ml of CuSO4 (r.o.) was added and heated to boiling (identification C). The melting point was also determined using the Buchi Melting Point B-454 instrument [9, 10]. 2.3 Identification and Determination According to Ph. Eur. 9 The chemical substance was identified according to Ph. Eur. 9, where the look of the solution and the pH value were determined. These tests were performed on a solution prepared by dissolving 2.5 mg of niacinamide in distilled water R, and made up to 50 ml with the same solvent. The appearance of the solution was determined by comparison with a reference solvent, and the pH value was determined using a pH meter. Niacinamide content is also determined according to Ph. Eur., whereby 0.250 g of niacinamide was dissolved in 20 ml of acetic acid, anhydrous R, heated and 5 ml of acetic acid, anhydride R was added. It was titrated with 0.1 M perchloric acid, with the indicator crystal violet, solution R, until the color changed in greenish blue. A blank and two parallel solutions of the substance were titrated [11, 12]. 2.4 Preparation of Solutions for the Characterization of Chemical Substance The niacinamide standard solution for the purpose of characterization was prepared by dissolving 10 mg of the substance in 100 ml of distilled water (c = 0.1 mg/ml), and this solution was diluted by measuring 4 ml of the basic solution into a 10 ml measuring vessel (c = 0.04 mg/ml). Series of 5 solutions of the standard substance of niacinamide and series of 10 solutions of the chemical substance niacinamide was prepared in the same way. The absorption spectrum of the base solution of the standard substance and the chemical substance was preliminarily recorded, in order to determine the optimal conditions for further tests. Analysis was carried out using a double beam UV-VIS spectrophotometer Shimatzu, UV-1800, using quartz cells and Shimatzu software at 258 nm, 260 nm, 262 nm, 264 nm and 266 nm. From results, standard deviation and relative standard deviation (RSD) were calculated [13].
446
M. Emira et al.
2.5 Preparation of Solutions for Determining the Linearity of the UV-VIS Method The stock solution was prepared by dissolving 10 mg of the niacinamide standard in a 100 ml measuring cup. Series of 5 solutions with concentrations of 0.005 mg/ml, 0.015 mg/ml, 0.03 mg/ml, 0.04 mg/ml, 0.05 mg/ml was made by diluting the appropriate volume of the stock solution with distilled water in individual measuring vessels of 10 ml. Linearity is expressed using the standard equation of direction, and confirmed using the correlation coefficient R2 .
3 Results and Discussion 3.1 Identification According to Ph. Yug. IV By identifying the chemical substance niacinamide according to Ph. Yug. IV we reached the following results: a) Solubility: – the substance is easily soluble in water: 1 g is dissolved in 4 ml of water; – the substance is easily soluble in ethanol: 1 g is dissolved in 9 ml of ethanol; – the substance is insoluble in ether: 1 g did not dissolve in 1000 ml of ether. b) Identification A: when heated, it melts and develops the smell of pirdin. c) Identification B: when mixed with 1M NaOH and heated to boiling, ammonia fumes are evolved. d) Identification C: when dissolved in water and mixed with CuSO4 , it turns blue. e) Melting point: 129 °C. 3.2 Identification and Determination According to Ph. Eur. 9 By identifying the chemical substance niacinamide according to Ph. Eur. 9, using solution S, we reached the following results: a) Look of the solution: solution S is clear and not more intensely colored than the reference solvent. b) pH value: solution S has a pH value of 6.0–7.5 (slightly acidic to neutral). For the purpose of determination, a blank sample and two parallel solutions of the chemical substance prepared in the previously described manner were titrated. The volume required for the titration of the niacinamide solution was calculated from the mean value of the volume of perchloric acid used. The results are shown in Table 1. Taking into account the fact that 1 ml of 0.1 M perchloric acid corresponds to 12.21 mg of niacinamide, we calculated that the chemical substance contains 99.39% of niacinamide, which is in accordance to the declared content.
Characterization of the Chemical Substance Niacinamide
447
Table 1. The volume of perchloric acid used to titrate the niacinamide solution #
The volume of perchloric acid (ml)
Blank sample
0
Sample 1
20.2
Sample 2
20.5
Mean value
20.35
3.3 Characterization of a Chemical Substance By recording the absorption spectrum of the solution of the standard and chemical substance niacinamide in the range 200–400 nm, the absorption maximum of 262 nm was determined. Further recordings were made at wavelengths near the absorption maximum: 258 nm, 260 nm, 262 nm, 264 nm and 266 nm. After recording, the mean value, standard deviation and relative standard deviation (RSD) were determined. The results are shown in Tables 2 and 3. Table 2. Results of the characterization of the niacinamide standard and the check standard Std/Abs 266 nm
264 nm
262 nm
260 nm
258 nm
1
0.763412
0.852264
0.905258
0.876038
0.841843
2
0.736550
0.852570
0.905182
0.875641
0.841919
3
0.763367
0.852066
0.905258
0.876160
0.841736
4
0.763779
0.852386
0.905487
0.876129
0.841644
5
0.763580
0.852280
0.905380
0.76022
0.841888
Avg
0.7635376 0.8523132
0.905313
0.875998
0.841806
Stdev
0.0001621 0.000184307 0.000120391 0.000207984 0.000114024
RSD
0.021
Check
266 nm
264 nm
1
0.766876
0.855560
0.022
0.013
0.024
0.014
262 nm
260 nm
258 nm
0.908859
0.879868
0.845261
2
0.767197
0.856003
0.908829
0.879883
0.845444
Avg
0.7670365
0.8557815
0.908844
0.8798755
0.8453525
f
101.53
101.59
101.60
101.55
101.57
The obtained results for the standard and chemical substance confirm the satisfactory value, expressed by the relative standard deviation (RSD ≤ 2%). 3.4 Linearity of the UV-VIS Method for the Determination of Niacinamide Linearity was determined on a series of 5 niacinamide standard solutions in the concentration range 0.005–0.05 mg/ml. Imaging was performed at wavelengths near the
448
M. Emira et al. Table 3. Results of the characterization of the chemical substance niacinamide #
mg
Abs (266 % nm)
Abs (264 % nm)
Abs (262 % nm)
1
10.3
0.783066
99.47 0.874863
99.56 0.929718
99.60
2
10.2
0.772583
99.10 0.862579
99.12 0.917038
99.21
3
10.2
0.772461
99.08 0.863022
99.17 0.917450
99.25
4
10.0
0.763733
99.92 0.852310
99.90 0.905319
99.90
5
10.2
0.774338
99.33 0.864700
99.36 0.918365
99.35
6
10.2
0.774353
99.33 0.864700
99.36 0.918564
99.37
7
10.4
0.786377
98.93 0.877945
99.95 0.932953
98.99
8
9.9
0.761551 100.65 0.850296 100.67 0.903458 100.70
9
10.0
0.762192
99.72 0.850677
99.71 0.903137
99.66
10
10.1
0.768539
99.56 0.857498
99.51 0.910736
99.50
Avg
10.15 0.771919
99.51 0.861859
99.53 0.929718
99.55
Stdev
0.48
0.46
0.45
RSD
0.48
0.47
0.46
%
Abs (258 nm)
μg/ml
#
mg
Abs (260 nm)
%
1
10.3
0.899734
99.62
0.863464
99.48
102.56
2
10.2
0.887177
99.19
0.851639
99.08
101.18
3
10.2
0.887085
99.18
0.851547
99.07
101.17
4
10.0
0.875778
99.87
0.841385
99.85
100.02
5
10.2
0.888245
99.31
0.853333
99.28
101.41
6
10.2
0.888626
99.35
0.853149
99.26
101.42
7
10.4
0.902832
99.00
0.866608
98.89
102.90
8
9.9
0.874161
100.70
0.839096
100.58
99.74
9
10.0
0.874268
99.70
0.839783
99.66
99.82
10
10.1
0.881302
99.51
0.846634
99.48
100.65
Avg
10.15
0.885920
99.54
0.850664
99.46
101.09
Stdev
0.46
0.46
1.04
RSD
0.46
0.47
1.03
absorption maximum: 258 nm, 260 nm, 262 nm, 264 nm and 266 nm. It is represented by a standard calibration curve (Figs. 2, 3, 4, 5 and 6) and confirmed by the correlation coefficient R2 . The results are shown below (Table 4).
Characterization of the Chemical Substance Niacinamide
449
Fig. 2. Calibration curve mg/ml-absorbance at the concentration range 0.005–0.05; λ = 258 nm
Fig. 3. Calibration curve mg/ml-absorbance at the concentration range 0.005–0.05; λ = 260 nm
The values of the correlation coefficient confirm the satisfactory linearity of the UVVIS method at all wavelengths at which the recording was performed (R2 ≥ 0.9995), and the method can be considered acceptable for further tests.
4 Conclusion The characterization of the chemical substance niacinamide was performed in relation to the certified standard substance, whereby the identity and declared content were confirmed. Confirmation of identity and determination of content were carried out according to the guidelines of Ph. Yug. IV and Ph. Eur. 9, and UV-VIS spectrophotometrically, whereby the chemical substance met all the set requirements. It can be concluded that the chemical substance has a satisfactory level of purity compared to the certified standard substance, and that as such it can be used for various purposes in the future, as a working standard. Also, the test confirmed the linearity of the UV-VIS method for the determining of niacinamide at selected wavelengths, and we conclude that this method
450
M. Emira et al.
Fig. 4. Calibration curve mg/ml-absorbance at the concentration range 0.005–0.05; λ = 262 nm
Fig. 5. Calibration curve mg/ml-absorbance at the concentration range 0.005–0.05; λ = 264 nm
Fig. 6. Calibration curve mg/ml-absorbance at the concentration range 0.005–0.05; λ = 266 nm
Characterization of the Chemical Substance Niacinamide
451
Table 4. Results of determining the linearity of the niacinamide solution at wavelengths of 258 nm, 260 nm, 262 nm, 264 nm and 266 nm c (mg/ml)
258 nm
260 nm
262 nm
264 nm
266 nm
0.005
0.115881
0.120482
0.124110
0.117337
0.105713
0.015
0.326529
0.340315
0.351146
0.330396
0.295807
0.03
0.656241
0.683180
0.704890
0.662823
0.593597
0.04
0.876913
0.912886
0.941997
0.886273
0.794479
0.05
1.094097
1.128096
1.164027
1.094556
0.980820
can be used for further tests, but only in monocomponent products with niacinamide. For products containing more active principles, prior separation of niacinamide is required, and the described method cannot be used for these purposes.
References 1. DrugBank. https://go.drugbank.com/drugs/DB02701. Last accessed 08 May 2023 2. Boo, Y.C.: Mechanistic basis and clinical evidence for the applications of nicotinamide (niacinamide) to control skin aging and pigmentation. Antioxidants 10, 1315 (2021) 3. Braidy, N., et al.: Role of nicotinamide adenine dinucleotide and related precursors as therapeutic targets for age-related degenerative diseases: rationale, biochemistry, pharmacokinetics, and outcomes. Antioxid. Redox Signal. 30, 251–294 (2019) 4. Zhen, A.X., et al.: Niacinamide protects skin cells from oxidative stress induced by particulate matter. Biomol. Ther. 27, 562–569 (2019) 5. Snaidr, V.A., Damian, D.L., Halliday, G.M.: Nicotinamide for photoprotection and skin cancer chemoprevention: a review of efficacy and safety. Exp. Dermatol. 28(1), 15–22 (2019) 6. Palazzo, L., Mikolcevic, P., Mikoc, A., Ahel, I.: ADP-ribosylation signalling and human disease. Open Biol. 9, 190041 (2019) 7. Larranaga, M.D., Lewis, R.J., Lewis, A.: Hawley’s Condensed Chemical Dictionary, 16th ed., p. 966. Wiley, Hoboken (2016) 8. PubChem. Open Chemistry Database. https://pubchem.ncbi.nlm.nih.gov/compound/936. Last accessed 08 May 2023 9. Daci´c, M., Sadadinovi´c, J.: Praktikum iz farmaceutske hemije I s teorijskim osnovama, pp. 169–170. Faculty of Pharmacy and Health, Travnik (2016) 10. Yugoslav pharmacopoeia IV (Ph Yug. IV) (1991) 11. Alagi´c-Džambi´c, L., Daci´c, M.: Praktikum iz farmaceutske hemije II i III s teorijskim osnovama, pp. 137–138. Faculty of Pharmacy and Health, Travnik (2020) 12. European Pharmacopoeia (Ph. Eur.), 9th ed. (2019) 13. Subhash Varun Kumar, V., Kavitha, J., Lakshami, K.S.: Spectrophotometric quantification of niacinamide in pharmaceutical dosage form by multivariate calibration technique. Res. J. Pharm. Technol. 14(4), 2013 (2021)
In Vitro Aerodynamic Comparison of Protective Masks ˇ car3 A. Uzunovi´c1 , E. Mlivo2 , M. Daci´c2(B) , S. Pilipovi´c1 , M. Šupuk1 , and H. Canˇ 1 Agency for Medicinal Products and Medical Devices, Titova 9, Sarajevo, Bosnia and
Herzegovina 2 Faculty of Pharm and Health, Slavka Gavranˇci´ca 17c, Travnik, Bosnia and Herzegovina
[email protected] 3 Faculty of Pharmacy, University of Sarajevo, Zmaja od Bosne 8, Sarajevo, Bosnia and
Herzegovina
Abstract. Most of the causative agents of respiratory infectious diseases are transmitted by droplets or through the air aerosols, and protective masks are recommended as a basic non-pharmaceutical intervention for prevention or slowing down their spread. In the conditions of the pandemic, there is a shortage of commercial masks and materials for their production, and an insufficient number of laboratories authorized to test protective equipment, have resulted in an increased need for new potential materials for making masks and alternative methods for their testing. In this research, a simple and fast method for testing the efficiency of particle filtration is proposed as the most important parameter of mask quality. The method is based on the passage of fluticasone propionate aerosol through the mask sample and then through the eight-stage Andersen cascade impactor in order to particles were classified on the basis of size, considering that the efficiency of filtration depends to the greatest extent on the size of the investigated particles. The filtration efficiency of the mask is determined by the mass of the particles that pass through the mask compared to the positive control. Masks with different properties and filtration mechanisms were tested: medical mask, filtering half mask and several different cloth face masks, and the results showed that the method is adequate for all tested types of protective masks. Although the filter half mask and medical mask showed the highest filtration efficiency for all particle sizes, multi-layer cloth masks can represent an equally good barrier to larger particles and significantly reduce the spread of droplet-borne viruses through and therefore as a last resort, they can represent a good alternative in cases of shortage of commercially available certified materials and masks. Keywords: Filtration efficiency · Mask · Respirator · Filtration · Andersen cascade impactor · HPLC
1 Introduction One of the basic non-pharmaceutical interventions in the fight against the spread of COVID-19 was the use of face masks and related interventions. Generally speaking, face masks, hand washing and social distancing are the most commonly applied measures in the fight against the transmission of respiratory infections [1, 2]. Masks are © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 A. Badnjevi´c and L. Gurbeta Pokvi´c (Eds.): MEDICON 2023/CMBEBIH 2023, IFMBE Proceedings 93, pp. 452–462, 2024. https://doi.org/10.1007/978-3-031-49062-0_49
In Vitro Aerodynamic Comparison of Protective Masks
453
classified into three basic categories: face masks, medical (surgical) masks and filter half masks (respirators). Face masks are masks that are made from commercially available double-layer, densely woven cotton materials [3]. Wearing this type of mask is considered to be a short-term and immediate practical solution for individuals living in developing countries who are trying to reduce their exposure to highly polluted air. They are very popular because they are cheap, easily available and washable [4–6]. Medical (surgical) masks are made of fabrics with a multi-layer structure consisting of a leak-proof layer, a high-density filter layer, and a layer that comes into direct contact with the skin [7]. The requirements for medical masks are specified in the standard EN 14683:2019+AC:2019. According to the above standard, medical masks are divided into two types (type I and type II) according to the efficiency of bacterial filtration, with type II being further divided depending on whether the mask is splash-proof or not. Type I medical masks should only be used by patients and other people to reduce the risk of spreading infections, especially in epidemics or pandemics and they are not intended for use by healthcare workers [8, 9]. Filter half-masks (respirators) are also made of several layers of fabric (often polypropylene); compared to medical masks, they have one more pre-filtration layer that is usually used for shaping, allowing a better fit to the face, they have a higher density and therefore represent a greater barrier to smaller particles [7]. In order to be placed on the market in Europe, they must undergo a procedure for assessing compliance with the European regulation EU2016/425. In addition, they must meet the requirements of the European standard EN 149:2001+A1:2009 [3]. The level of protection offered by filter half-masks and face masks is defined by the percentage of particles from the environment that penetrate inside the protective device, where there are two basic penetration paths: leakage through the seals and through the filter medium [10–12]. The COVID-19 pandemic has shown that it is in such and similar situations more than ever, raising global awareness and knowledge about the importance of respecting the essential requirements is needed to guarantee the appropriate quality, performance and safety of medical products [13]. Although the regulatory framework prescribes post-market surveillance of medical devices, the process itself is not harmonized with international standards. The lack of harmonization within post-market surveillance results in an environment of increased adverse events involving medical devices and overall mistrust in medical device diagnosis and treatment results [14]. The filtration efficiency of a certain type of mask is tested, depending on the individual type, according to the previously mentioned standards. Also, different requirements are set depending on whether they are face masks, medical masks or respirators and depending on the basic characteristics of certain type of masks [15, 16].
2 Materials and Methods 2.1 Chemicals, Apparatus and Samples The following chemicals were used for the research: sodium dodecyl sulfate (≥99%, Carl Roth GmbH, Germany), glacial acetic acid (≥99.8%, Honeywell GmbH, Germany), methanol and ethanol of HPLC purity (≥99.9%, Fisher Scientific U.K., United Kingdom), fluticasone propionate reference standard (99.5%, Ph.Eur.) and fluticasone propionate inhalation suspension (50 µg and 125 µg). The following solutions were
454
A. Uzunovi´c et al.
used in the analysis: buffer, 0.01 M sodium dodecyl sulfate and 0.1% glacial acetic acid; solution A: methanol and buffer (20:80). The following mobile phase was used: acetonitrile and solution A (50:50) diluent: methanol and water (60:40). For the testing, eight-stage Andersen cascade impactor (Fig. 1) EP App D/USP App 1 and 3 (Westech Scientific Instruments, United Kingdom), HPLC with Photodiode Array (PDA) detector (Shimadzu Corporation, Japan) and HPLC column Zorbax SB-C18 50 mm × 4.6 mm, 1.8 µm (Agilent Technologies Inc., USA) was used. The samples used for testing were cut from finished products or from raw materials for making masks (Fig. 2) and their characteristics are given in the Table 1.
Fig. 1. Schematic representation of the cascade impactor
2.2 Principle and Procedure The sample of the mask material is placed on the inlet opening of the eight-stage cascade impactor, into which an aerosol of fluticasone propionate is introduced, which passes through the mask material and the impactor under vacuum. Based on the mass of particles that pass through the mask and their distribution in the sections of the cascade impactor, as a percentage compared to the positive control, the filtration efficiency of the mask is determined. First, a positive control without a sample was performed. Clean plates were
In Vitro Aerodynamic Comparison of Protective Masks
455
Table 1. Characteristics of tested samples Sample
Abbreviation Number of layers Thickness (mm) Mass per area unit (g/m2 )
Medical mask
Medical
3
0.38
81
Polyester mask
PES 3x
3
1.09
621
Polyester mask
PES 2x
2
0.81
404
Polyester viroblock PES VB mask
3
1.10
597
Cotton mask
Cotton 1x
1
0.39
140
Cotton mask
Cotton 2x
2
0.77
280
FFP2 filter mask
FFP2
4
0.87
191
Fig. 2. Raw materials for making masks (A - Medical mask; B - FFP2 filter mask; C - Cotton mask; D - Polyester mask 3x; E - Polyester Viroblock mask 3x, F - Polyester mask 2x)
then placed in the impactor, and the mask sample was placed between the first section and the inlet opening (cone) and the same procedure was repeated. After removing the cascade impactor, the sample mask, inlet port, plates and filter were washed with diluent, and the associated fractions were collected in measuring vessels of the appropriate volume and diluted. A standard solution was prepared (0.8 µg/ml fluticasone propionate in diluent); standard solution and test solution are analyzed chromatographically, at a mobile phase
456
A. Uzunovi´c et al. Table 2. Preparation of test solutions Fluticasone propionate 125 µg
Fluticasone propionate 50 µg
Measuring cup (ml)
Dilution factor
Measuring cup (ml)
Dilution factor
Mask
200
0.002
100
0.01
Entry port
200
0.002
100
0.01
Level 0
50
0.02
50
0.02
Level 1
25
0.04
25
0.04
Level 2
25
0.04
25
0.04
Fraction
Level 3
10
0.01
50
0.02
Level 4
100
0.01
50
0.02
Level 5
100
0.01
50
0.02
Level 6
25
0.04
25
0.04
Level 7
25
0.04
25
0.04
Filter
25
0.04
25
0.04
flow of 2 ml/min, an injection volume of 50 µl and a temperature of 40 °C; detection was performed at 239 nm (Table 2).
3 Results and Discussion 3.1 Reference Standards The chromatogram of the standard solution is shown in Fig. 3, and the response factors benefits of the system in Table 3. Table 3. Standards response factors Fluticasone propionate (99.5%)
Peak area
Response factor (RF)
Standard 1 (m = 1.651 mg)
63167
1.30032E-08
63067
1.30238E-08
63135
1.30098E-08
59991
1.3136E-08
60046
1.31239E-08
60078
1.31169E-08
Standard 2 (m = 1.584 mg)
Mean
1.30689E-08
SD
6.27415E-11
RSD (%)
0.48
In Vitro Aerodynamic Comparison of Protective Masks
457
Fig. 3. The chromatogram of the standard solution
3.2 Positive Control The positive control was performed by releasing the test aerosol through the cascade impactor without a mask sample. The 125 µg fluticasone propionate inhaler had a MMAD (Mass Median Aerodynamic Diameter) = 4.05 µm and GSD (geometric standard deviation) = 1.61, while for the inhaler with 50 µg MMAD = 3.61 µm and GSD = 1.47. For more details see Tables 4 and 5. Table 4. Positive control: the results based on the individual stages and the final filter #
Test 1
Test 2
Test3
Peak area mcg/activation Peak area mcg/activation Peak area mcg/activation Entry port 158265
20.684
316987
41.427
108789
71.088
18540
1.211
56623
3.700
Level 0
9159
0.598
Level 1
26826
0.876
53340
1.743
141448
4.621
Level 2
56424
1.844
112399
3.672
203849
6.660
Level 3
70250
4.590
140251
9.165
116255
15.193
Level 4
54456
3.558
108748
7.106
69985
9.146
Level 5
19235
1.257
38448
2.512
20051
2.620
Level 6
11100
0.363
22589
0.738
15007
0.490
Level 7
4232
0.138
8429
0.275
7368
0.241
Filter
3736
0.122
7810
0.255
6429
0.210
458
A. Uzunovi´c et al. Table 5. Positive control: the results based on the particle size distribution
Particle size (µm)
Test 1 (µg)
Test 2 (µg)
Test 3 (µg)
0–9
12.748
25.466
39.181
≥5
2.080
4.141
9.124
5 µm) and via aerosols (particles < 5 µm), while the size of the viruses themselves is around 20–1000 nm [19]. That is why it is in operation, except for the total efficiency of filtering and distribution by sections of the cascade impactor, efficiency was also considered particle filtering: ≤1 µm; 0–5 µm; 5–9 µm; 3 ± 0.3 µm. It should be emphasized that the efficiency of the filtration of the protective material is tested with the proposed method mask, not taking into account permeability when wearing it. Although standards for medical masks (EN 14683:2019, ASTM F2100) only define requirements for filter material, standards for respirators set requirements for overall mask throughput that includes system throughput for fastening, permeability of the exhalation valve (if installed) and permeability of the filter materials itselfs (EN 149:2001, NIOSH: 42 CFR 84).
In Vitro Aerodynamic Comparison of Protective Masks Table 6. Efficiency of filtration by cascade impactor sections #
Medical (%)
PES 3x (%)
PES 2x (%)
PES VB (%)
Level 1
99.57
95.04
93.20
97.24
Level 2
98.98
98.37
92.69
97.80
Level 3
99.30
98.21
89.30
96.11
Level 4
99.14
97.12
79.08
90.62
Level 5
98.49
93.16
64.72
78.93
Level 6
97.70
84.44
48.46
63.39
Level 7
94.45
74.94
34.84
47.47
Filter
76.55
59.24
48.91
33.81
#
Cotton 1x
Cotton 2x
FFP2
Level 1
75.81
93.92
99.38
Level 2
73.59
92.39
99.72
Level 3
64.90
89.31
100.00
Level 4
39.62
88.98
100.00
Level 5
40.42
82.80
100.00
Level 6
36.38
66.33
99.44
Level 7
20.73
51.93
95.99
Filter
14.67
34.60
93.75
459
460
A. Uzunovi´c et al. Table 7. Filtration efficiency of particles in the size range of interest
Type of mask
≥5 µm
≥5 µm (RSD %)