137 88 94MB
English Pages 543 [533] Year 2023
Sachi Nandan Mohanty Vicente Garcia Diaz G. A. E. Satish Kumar (Eds.)
470
Intelligent Systems and Machine Learning First EAI International Conference, ICISML 2022 Hyderabad, India, December 16–17, 2022 Proceedings, Part I
Part 1
123
Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering Editorial Board Members Ozgur Akan, Middle East Technical University, Ankara, Türkiye Paolo Bellavista, University of Bologna, Bologna, Italy Jiannong Cao, Hong Kong Polytechnic University, Hong Kong, China Geoffrey Coulson, Lancaster University, Lancaster, UK Falko Dressler, University of Erlangen, Erlangen, Germany Domenico Ferrari, Università Cattolica Piacenza, Piacenza, Italy Mario Gerla, UCLA, Los Angeles, USA Hisashi Kobayashi, Princeton University, Princeton, USA Sergio Palazzo, University of Catania, Catania, Italy Sartaj Sahni, University of Florida, Gainesville, USA Xuemin Shen , University of Waterloo, Waterloo, Canada Mircea Stan, University of Virginia, Charlottesville, USA Xiaohua Jia, City University of Hong Kong, Kowloon, Hong Kong Albert Y. Zomaya, University of Sydney, Sydney, Australia
470
The LNICST series publishes ICST’s conferences, symposia and workshops. LNICST reports state-of-the-art results in areas related to the scope of the Institute. The type of material published includes • Proceedings (published in time for the respective event) • Other edited monographs (such as project reports or invited volumes) LNICST topics span the following areas: • • • • • • • •
General Computer Science E-Economy E-Medicine Knowledge Management Multimedia Operations, Management and Policy Social Informatics Systems
Sachi Nandan Mohanty · Vicente Garcia Diaz · G. A. E. Satish Kumar Editors
Intelligent Systems and Machine Learning First EAI International Conference, ICISML 2022 Hyderabad, India, December 16–17, 2022 Proceedings, Part I
Editors Sachi Nandan Mohanty VIT-AP University Amr¯avati, Andhra Pradesh, India
Vicente Garcia Diaz University of Oviedo Oviedo, Spain
G. A. E. Satish Kumar Vardhaman College of Engineering Hyderabad, India
ISSN 1867-8211 ISSN 1867-822X (electronic) Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering ISBN 978-3-031-35077-1 ISBN 978-3-031-35078-8 (eBook) https://doi.org/10.1007/978-3-031-35078-8 © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2023 This work is subject to copyright. All rights are reserved 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
Preface
We are delighted to introduce the proceedings of the first edition of the 2022 European Alliance for Innovation (EAI) International Conference on Intelligent Systems and Machine Learning (ICISML 2022). ICISML 2022 is the premier conference in the field of emerging technologies. In today’s world, intelligent systems are a new and rapidly growing technology. A large volume of data can be generated and made available for analysis. Humans lack the cognitive ability to comprehend such massive volumes of data. Humans can use machine learning (ML) to process enormous volumes of data and get insights into how the data behaves. These machines can learn from their mistakes and do human-like tasks. The 75 papers presented were selected from 209 submissions. The conference has four different technical tracks: Track 1: Image Recognition, Track 2: Machine Learning, Track 3: Intelligent Systems and Machine Learning Applications, and Track 4: Intelligent Communications Networks. The technical programme also featured four keynote talks and one technical workshop. The title of the technical workshops organised in collaboration with the software industry, “CORTEVA Agriscience,” has the theme “Sequences to Satellites: Applications of Artificial Intelligence and Machine Learning in the Agriculture Domain.” The average number of papers per reviewer was three. Around 192 technical experts participated in the review process from across the globe. Three experts in various fields reviewed each paper. A technical programme committee member finally took the decision and communicated it to the authors as per reviewer reports. Coordination with the general chair, Vicente Garcia Diaz, University of Oviedo, Spain, was essential for the success of the conference. We sincerely appreciate their constant support and guidance. It was also a great pleasure to work with such an excellent organising committee team and we appreciate their hard work in organising and supporting the conference. We strongly believe that the ICISML 2022 conference provided a good forum for all researchers, developers, and practitioners to discuss all application aspects of science and technology that are relevant to smart grids. We also expect that future ICISML conferences will be as successful and stimulating as indicated by the contributions presented in this volume. July 2023
Sachi Nandan Mohanty Vicente Garcia Diaz G. A. E. Satish Kumar
Conference Organization
Steering Committee Imrich Chlamtac T. Vijender Reddy (Chairman) M. Rajasekhar Reddy (Vice Chairman) T. Upender Reddy E. Prabhakar Reddy J. V. R. Ravindra
Bruno Kessler Professor, University of Trento, Italy Vardhaman College of Engineering, Hyderabad, India Vardhaman College of Engineering, Hyderabad, India Vardhaman College of Engineering, Hyderabad, India Vardhaman College of Engineering, Hyderabad, India Vardhaman College of Engineering, Hyderabad, India
Organizing Committee General Chair Vicente Garcia Diaz
University of Oviedo, Spain
General Co-chairs G. A. E. Satish Kumar Milos Stojmenovic George Tsaramirsis Sachi Nandan Mohanty
Vardhaman College of Engineering, Hyderabad, India Singidunum University, Serbia Higher College of Technology, Abu Dhabi, UAE Singidunum University, Serbia
Technical Program Chairs G. A. E. Satish Kumar Tanupriya Choudhury
Vardhaman College of Engineering, Hyderabad, India University of Petroleum & Energy Studies, Dehradun, India
viii
Conference Organization
Sachi Nandan Mohanty Sven Groppe
Singidunum University, Serbia University of Lübeck, Germany
Technical Program Co-chairs Mohammed Altaf Ahmed Adil O. Khadidos Sanjay Misra Maria Spichkova
Prince Sattam Bin Abdulaziz University, Saudi Arabia King Abdulaziz University, Jeddah, Saudi Arabia Covenant University, Nigeria RMIT University, Australia
Web Chair A. Pramod Kumar
Vardhaman College of Engineering, India
Publicity and Social Media Chairs Sarika Jain Nonita Sharma Monika Mangala Suvendu Kumar Pani
NIT Kurukshetra, India Indira Gandhi Delhi Technical University for Women, New Delhi, India Dwarkadas J. Sanghvi College of Engineering Mumbai, India Krupajala Engineering College, Bhubaneswar, India
Organizing Chair A. Pramod Kumar
Vardhaman College of Engineering, Hyderabad, India
Convenors M. Naresh Kumar D. Krishna
Vardhaman College of Engineering, Hyderabad, India Vardhaman College of Engineering, Hyderabad, India
Special Session Chairs Sarita Mohanty Tejaswini Kar
CPGS, Odisha University Agriculture University, Bhubaneswar, India KIIT, Bhubaneswar, India
Conference Organization
Website Chair D. Praveen Kumar
Vardhaman College of Engineering, Hyderabad, India
Registration Chairs Ch. Sulakshana Sangeeta Singh
Vardhaman College of Engineering, Hyderabad, India Vardhaman College of Engineering, Hyderabad, India
Finance Chairs J. Krishna Chaithanya B. Srikanth
Vardhaman College of Engineering, Hyderabad, India Vardhaman College of Engineering, Hyderabad, India
Workshops Chair Nonita Sharma
NIT Jalandhar, India
Sponsorship and Exhibits Chairs Jatin Kumar Das I. Babu M. Nagarjuna
SRM-AP University, Andhra Pradesh, India Vardhaman College of Engineering, Hyderabad, India Vardhaman College of Engineering, Hyderabad, India
Publication Chair Suneeta Satpathy Priya Gupta
Sri Sri University Odisha, India JNU New Delhi, India
Panel Chair Mohammed Altaf Ahmed
Prince Sattam Bin Abdulaziz University, Saudi Arabia
ix
x
Conference Organization
Local Chair J. Krishna Chaitanya
Vardhaman College of Engineering, Hyderabad, India
Technical Program Committee Sanjay Misra Milos Stojmenovi Georgios Tsaramirsis Sven Groppe Mohammad Yamin Mukesh Misra Neha Mohanty Aleksandar Jevremovic Jinghua Groppe Abdulrhman M. Alshareef Gourav Sen Gupta S. K. Ramesh María José Navarro Avilés Ishani PriyaDarshini Laxmidhar Behera (Director) Santunu Choudhury Durga Prasad Mohapatro Debasis Samanta E. Laxmi Lydia Mukta Jagdish Mahendran Arumugam Mrutyunjaya Panda Abdulsattar Abdullah Hamad Bhagirathi Nayak Bunil Kumar Balabantaray Priya Gupta Martine Gadille Deepak Gupta
Covenant University, Nigeria Singidunum University, Serbia Higher Colleges of Technology Women’s Campus, Abu Dhabi, UAE University of Lübeck, Germany King Abdulaziz University, Saudi Arabia Massey University, New Zealand New Jersey Institute of Technology, USA University of Lübeck, Germany University of Lübeck, Germany King Abdulaziz University, Saudi Arabia Massey University, New Zealand California State University, USA Gijon Hospital, Spain UC Berkeley, USA Indian Institute of Technology Mandi, India IIT Jodhpur, India NIT Rourkela, India IIT Kharagpur, India Vignan’s Institute of Information Technology, Visakhapatnam, India Vardhaman College of Engineering, Hyderabad, India IIT Saveetha, India Utkal University, India Hamad Imam University College, Iraq Sri Sri University, India National Institute of Technology, Meghalaya, India Jawaharlal Nehru University, India AMU, France National Institute of Technology, Arunachal Pradesh, India
Conference Organization
Gouse Baig Sanjay Kuamr Panda Deepak Gupta
xi
Vardhaman College of Engineering, Hyderabad, India NIT Warangal, India Maharaja Agrasen Institute of Technology, India
Contents – Part I
Intelligent Systems and Machine Learning Applications in Health Care Improving Multi-class Brain Tumor Detection Using Vision Transformer as Feature Extractor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Adeel Ahmed Abbasi, Lal Hussain, and Bilal Ahmed Measles Rash Disease Classification Based on Various CNN Classifiers . . . . . . . Lohitha Rani Chintalapati, Trilok Sai Charan Tunuguntla, Yagnesh Challagundla, Sachi Nandan Mohanty, and S. V. Sudha Brain Imaging Tool in Patients with Trans Ischemic Attack: A Comparative Research Study Analysis of Computed Tomography and Magnetic Resonance Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . R. Bhuvana and R. J. Hemalatha EEG-Based Stress Detection Using K-Means Clustering Method . . . . . . . . . . . . . Soumya Samarpita and Rabinarayan Satpathy Detection of Psychological Stability Status Using Machine Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Manoranjan Dash, M. Narayana, Nampelly Pavan Kalyan, Md Azam Pasha, and D. Chandraprakash GLCM Based Feature Extraction and Medical X-ray Image Classification Using Machine Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jyotiranjan Rout, Swagat Kumar Das, Priyabrata Mohalik, Subhashree Mohanty, Chandan Kumar Mohanty, and Susil Kumar Behera
3
15
28
35
44
52
Multi-filter Wrapper Enhanced Machine Learning Model for Cancer Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bibhuprasad Sahu and Sujata Dash
64
An Interactive Web Solution for Electronic Health Records Segmentation and Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sudeep Mathew, Mithun Dolthody Jayaprakash, and Rashmi Agarwal
79
A Convolutional Neural Network Based Prediction Model for Classification of Skin Cancer Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vanshika Saini, Neelanjana Rai, Nonita Sharma, and Virendra Kumar Shrivastava
92
xiv
Contents – Part I
Multimodal Biomedical Image Fusion Techniques in Transform and Spatial Domain: An Inclusive Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 Nitin S. Thakare and Mukesh Yadav Early Prediction of Coronary Heart Disease Using the Boruta Method . . . . . . . . . 119 Vaibhav Satija, Mohaneesh Raj Pradhan, and Princy Randhawa Design and Implementation of Obesity Healthcare System (OHS) Using Flutter Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 Mehnaz Hikman Ud Din and Samar Mouti A Survey on Covid-19 Knowledge Graphs and Their Data Sources . . . . . . . . . . . 142 Hanieh Khorashadizadeh, Sanju Tiwari, and Sven Groppe Identify Melanoma Using CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 G. M. Shashidhara, Rashmi Agarwal, and Jitendra Suryavamshi Digital Forensic and Network Security Machine Learning Based Malware Analysis in Digital Forensic with IoT Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 Sreenidhi Ganachari, Pramodini Nandigam, Anchal Daga, Sachi Nandan Mohanty, and S. V. Sudha Malicious Codes Detection: Deep Learning Techniques . . . . . . . . . . . . . . . . . . . . . 184 Jasleen Gill and Rajesh Dhakad Securing Outsourced Personal Health Records on Cloud Using Encryption Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 Abhijeet Borade and Rashmi Agarwal Design and Evaluation Decentralized Transactional Network Based Blockchain Technology Using Omnet++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 Morched Derbali Movie Synchronization System Using Web Socket Based Protocol . . . . . . . . . . . 222 Amar Shukla, Thipendra Pal Singh, Vikas Mishra, Garima Goyal, Ishita Kanwar, Gauraang Sharma, and Tanupriya Choudhury A Study on Android Malware Detection Using Machine Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232 K. S. Ujjwal Reddy, S. Sibi Chakkaravarthy, M. Gopinath, and Aditya Mitra
Contents – Part I
xv
Intelligent Communication Wireless Networks A Survey on Deep Recurrent Q Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 M. V. K. Gayatri Shivani, S. P. V. Subba Rao, and C. N. Sujatha Predicting Credit Card Defaults with Machine Learning Algorithm Using Customer Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 Anushka, Nidhi Agarwal, Devendra K. Tayal, Vrinda Abrol, Deepakshi, Yashica Garg, and Anjali Jha Enhancement of Signal to Noise Ratio for QAM Signal in Noisy Channel . . . . . 278 Ali Salah Mahdi GSM Enabled Patient Monitoring System Using Arduino Application for Cardiac Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288 Samson Jebakumar, R. J. Hemalatha, and R. Kishore Kanna Reso-Net: Generic Image Resolution Enhancement Using Convolutional Autoencoders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298 Koustav Dutta and Priya Gupta Smart Traffic System with Green Time Optimization Using Fuzzy Logic . . . . . . 309 A. K. Kavin, R. Radha, Vishnu Prasad, and Bharathwaj Murali Early Diagnosis of Rheumatoid Arthritis of the Wrist Using Power Doppler Ultrasound: A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320 D. Priscilla Sharlet Asha and R. J. Hemalatha An Intrusion Detection System and Attack Intension Used in Network Forensic Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334 Saswati Chatterjee, Lal Mohan Pattnaik, and Suneeta Satpathy Comparison of Advanced Encryption Standard Variants Targeted at FPGA Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346 Nithin Shyam Soundararajan and K. Paldurai Characteristics and Analysis of ElectroGastroGram Signal . . . . . . . . . . . . . . . . . . 356 R. Chandrasekaran, S. Vijayaraj, and G. R. Jothi Lakshmi A Comparative Study of Power Optimization Using Leakage Reduction Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362 Pramod Kumar Aylapogu, A. Jayalakshmi, Hirald Dwaraka Praveena, and B. Kalivaraprasad
xvi
Contents – Part I
Design and Implementation of 4-bit High Speed Array Multiplier for Image Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382 Pramod Kumar Aylapogu, Kalivaraprasad Badita, P. Geetha, and Namratha Sama Internet of Things (IoT) Applications Smart Traffic Police Helmet: Using Image Processing and IoT . . . . . . . . . . . . . . . 393 Shoaib Hafeez, Ramesh Karnati, and Muni Sekhar Velpuru Interconnected Hospitals Using IOT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401 Subhra Debdas, Prem Bdr Shah, Maddhuja Sen, Priti Priya Das, Abhiyankar Shakti, and D. Venkat Prasad Varma Automatic Oxygen Ventilation and Monitoring System Using IoT . . . . . . . . . . . . 412 Madhunala Srilatha, K. Vinay, and Polemoni Jevardhan Raju Social Informatics Social Media Sentiment Analysis Using Deep Learning Approach . . . . . . . . . . . . 431 M. Mohamed Iqbal, K. S. Arikumar, Balaji Vijayan Venkateswaralu, and S. Aarif Ahamed Movie Recommendation Using Content-Based and Collaborative Filtering Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439 Anjali Jha, Nidhi Agarwal, Devendra K. Tayal, Vrinda Abrol, Deepakshi, Yashica Garg, and Anushka A Systematic Review on Recommender System Models, Challenges, Domains and Its Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451 Rajesh Garapati and Mehfooza Munavar Basha Movie Recommendation System Using Composite Ranking . . . . . . . . . . . . . . . . . 468 Aashal Kamdar and Irish Mehta Social Distancing and Face Mask Detection Using Open CV . . . . . . . . . . . . . . . . . 488 Majji Ramachandro, Ala Rajitha, Dasari Madhavi, Jagini Naga Padmaja, and Ganesh B. Regulwar Predicting the Likeliest Customers; Minimizing Losses on Product Trials Using Business Analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499 Tushar Nigam and Rashmi Agarwal
Contents – Part I
xvii
Television Price Prediction Based on Features with Machine Learning . . . . . . . . 507 Marumoju Dheeraj, Manan Pathak, G. R. Anil, and Mohamed Sirajudeen Yoosuf Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
Contents – Part II
Emerging Applications A Model for Engineering, Procurement, and Construction (EPC) Organizations Using Vendor Performance Rating System . . . . . . . . . . . . . . . . . . . . Sujit Kumar Panda, Sukanta Kumar Baral, Richa Goel, and Tilottama Singh F2PMSMD: Design of a Fusion Model to Identify Fake Profiles from Multimodal Social Media Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bhrugumalla L. V. S. Aditya, Gnanajeyaraman Rajaram, Shreyas Rajendra Hole, and Sachi Nandan Mohanty A Novel Model to Predict the Whack of Pandemics on the International Rankings of Academia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nidhi Agarwal and Devendra K. Tayal Credit Risk Assessment - A Machine Learning Approach . . . . . . . . . . . . . . . . . . . Thumpala Archana Acharya and Pedagadi Veda Upasan Development of Analytical DataMart and Data Pipeline for Recruitment Analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ashish Chandra Jha, Sanjeev Kumar Jha, and J. B. Simha
3
13
24
39
55
Data Homogeneity Dependent Topic Modeling for Information Retrieval . . . . . . Keerthana Sureshbabu Kashi, Abigail A. Antenor, Gabriel Isaac L. Ramolete, and Adrienne Heinrich
66
Pattern Discovery and Forecasting of Attrition Using Time Series Analysis . . . . Saumyadip Sarkar and Rashmi Agarwal
85
Resume Shortlisting and Ranking with Transformers . . . . . . . . . . . . . . . . . . . . . . . Vinaya James, Akshay Kulkarni, and Rashmi Agarwal
99
Hybrid Deep Learning Based Model on Sentiment Analysis of Peer Reviews on Scientific Papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 Ritika Sarkar, Prakriti Singh, Mustafa Musa Jaber, Shreya Nandan, Shruti Mishra, Sandeep Kumar Satapathy, and Chinmaya Ranjan Pattnaik
xx
Contents – Part II
Artificial Intelligence Based Soilless Agriculture System Using Automatic Hydroponics Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Aveen Uthman Hassan, Mu’azu Jibrin Musa, Yahaya Otuoze Salihu, Abubakar Abisetu Oremeyi, and Fatima Ashafa Mining Ancient Medicine Texts Towards an Ontology of Remedies – A Semi-automatic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 João Nunes, Orlando Belo, and Anabela Barros A Novel Oversampling Technique for Imbalanced Credit Scoring Datasets . . . . . 147 Sudhansu Ranjan Lenka, Sukant Kishoro Bisoy, Rojalina Priyadarshini, and Jhalak Hota A Blockchain Enabled Medical Tourism Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . 158 Nihar Ranjan Pradhan, Hitesh Kumar Sharma, Tanupriya Choudhury, Anurag Mor, and Shlok Mohanty Measuring the Impact of Oil Revenues on Government Debt in Selected Countries by Using ARDL Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 Mustafa Kamil Rasheed, Ahmed Hadi Salman, and Amer Sami Mounir Diagnosis of Plant Diseases by Image Processing Model for Sustainable Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181 Sasmita Pani, Jyotiranjan Rout, Zeenat Afroz, Madhusmita Dey, Mahesh Kumar Sahoo, and Amar Kumar Das Face Mask Detection: An Application of Artificial Intelligence . . . . . . . . . . . . . . . 193 Poonam Mittal, Ashlesha Gupta, Bhawani Sankar Panigrahi, Ruqqaiya Begum, and Sanjay Kumar Sen A Critical Review of Faults in Cloud Computing: Types, Detection, and Mitigation Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 Ramandeep Kaur and V. Revathi Video Content Analysis Using Deep Learning Methods . . . . . . . . . . . . . . . . . . . . . 222 Gara Kiran Kumar and Athota Kavitha Prediction of Cochlear Disorders Using Face Tilt Estimation and Audiology Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 Sneha Shankar, Sujay Doshi, and G. Suganya Quantum Data Management and Quantum Machine Learning for Data Management: State-of-the-Art and Open Challenges . . . . . . . . . . . . . . . . . . . . . . . . 252 Sven Groppe, Jinghua Groppe, Umut Çalıkyılmaz, Tobias Winker, and Le Gruenwal
Contents – Part II
xxi
Multivariate Analysis and Comparison of Machine Learning Algorithms: A Case Study of Cereals of America . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 Rashika Gupta, E. Lavanya, Nonita Sharma, and Monika Mangla Competitive Programming Vestige Using Machine Learning . . . . . . . . . . . . . . . . . 272 Ajay Dharmarajula, Challa Sahithi, and G. S. Prasada Reddy Machine Learning Techniques for Aspect Analysis of Employee Attrition . . . . . 286 Anamika Hooda, Purva Garg, Nonita Sharma, and Monika Mangla AI-Enabled Automation Solution for Utilization Management in Healthcare Insurance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297 Gaurav Karki, Jay Bharateesh Simha, and Rashmi Agarwal Real-Time Identification of Medical Equipment Using Deep CNN and Computer Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311 Jaya Rubi, R. J. Hemalatha, and Bethanney Janney Design of a Intelligent Crutch Tool for Elders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320 A. Josephin Arockia Dhivya and R. J. Hemalatha An Approach to New Technical Solutions in Resource Allocation Based on Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325 Tung Nguyen Trong, Nguyen Hai Vinh Cuong, Tran-Vu Pham, Nguyen Ha Huy Cuong, and Bui Thanh Khiet Gesture Controlled Power Window Using Deep Learning . . . . . . . . . . . . . . . . . . . 335 Jatin Rane and Suhas Mohite Novel Deep Learning Techniques to Design the Model and Predict Facial Expression, Gender, and Age Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 N. Sujata Gupta, Saroja Kumar Rout, Viyyapu Lokeshwari Vinya, Koti Tejasvi, and Bhargavi Rani A Comprehensive Review on Various Data Science Technologies Used for Enhancing the Quality of Education Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 355 Olfat M. Mirza AI/ML Based Sensitive Data Discovery and Classification of Unstructured Data Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 Shravani Ponde, Akshay Kulkarni, and Rashmi Agarwal Bias Analysis in Stable Diffusion and MidJourney Models . . . . . . . . . . . . . . . . . . 378 Luka Aniˇcin and Miloš Stojmenovi´c
xxii
Contents – Part II
Machine Learning Based Spectrum Sensing for Secure Data Transmission Using Cuckoo Search Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389 E. V. Vijay and K. Aparna Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
Intelligent Systems and Machine Learning Applications in Health Care
Improving Multi-class Brain Tumor Detection Using Vision Transformer as Feature Extractor Adeel Ahmed Abbasi1 , Lal Hussain2(B) , and Bilal Ahmed1 1 School of Computer Science and Engineering, Central South University, 932 Lushan S Rd,
Yuelu District, Changsha, Hunan, China 2 Department of Computer Science and IT, The University of Azad Jammu and Kashmir,
Neelum Campus, Muzaffarabad, Azad Kashmir, Pakistan [email protected]
Abstract. Accurate brain tumor subtypes classification is significant for prognosis and treatment. The aim of this research is to improve the multiclass brain tumor classification using vision transformer as feature extractor. In this study, we first optimized and employed deep learning ResNet101 for feature extraction and fed to machine learning classifiers for multi-class classification. We then optimized and employed vision transformer and fed these features to machine learning decision classifier. We measured the performance with standard performance metrics. The Artificial Intelligence vision transformer with decision tree classifier yielded highest multi-class classification performance with 99.89% accuracy and 1.00 AUC to detect pituitary followed by 97.69% accuracy and AUC of 0.96 to detect meningioma. The results are compared with ResNet101 with transfer learning. ResNet101 deep features by utilizing KNN yielded detection of pituitary tumor (98.80% accuracy, 0.99 AUC), glioma (93.47% accuracy, 0.93 AUC). The results revealed that proposed approach with vision transformer and decision tree features extractor are more robust in detecting multiclass brain tumor prediction. The proposed approach can be better utilized for betterment of treatment and prognosis to obtain improved clinical outcomes. Keywords: deep learning Brain tumor · types · convolution neural network · vision transformer
1 Introduction The brain tumor is one of the most commonly diagnosed diseases in humans without discrimination of any age group and sex. Generally, a brain tumor is separated into various stages (1, 2, 3.. etc.), a higher stage means higher severeness of the tumor, and earlier stage tumors have more probability of recovery. So, diagnosing the tumor initially is very important in order to initiate a fruitful treatment [1]. However, at earlier stages tumors are so small to detect for radiologists even from the quality magnetic resonance imaging (MRI). MRI technology gives more in-depth information about the tumor and is more efficient compared to old techniques. Thus, the chance of human error and competence © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2023 Published by Springer Nature Switzerland AG 2023. All Rights Reserved S. Nandan Mohanty et al. (Eds.): ICISML 2022, LNICST 470, pp. 3–14, 2023. https://doi.org/10.1007/978-3-031-35078-8_1
4
A. A. Abbasi et al.
(knowledge and health issues, etc.) compels researchers to take advantage of computers. Various computer-aided studies are suggested in the past to assist radiologists to detect brain tumors from MRI images [2–6]. Few researchers used Machine learning-based methods for classification purposes. Most of the machine learning-based methods are based on handcrafted features [7–12]. The key part of Machine Learning based classification tasks is features extraction which produces discriminant information of an image. Feature extraction techniques such as Scale-invariant Fourier transform (SIFT), Morphology, Texture, entropy and elliptic Fourier descriptors (EFDs) can be employed to extract features from medical raw images [12]. In the past, these methods were used separately and in a hybrid way along with various Machine Learning (ML) classification algorithms. For brain tumor classification there are a variety of Machine learning classification algorithms were reported [13–27]. Other studies such as [28] performed some morphological operations on the segmented part of an image and employed features extraction technique Space Pyramid Matching (SPM). Then, for the classification task, features were passed to ML classifiers such as Support vector Machine, (SVM), K-nearest neighbor (KNN), and Sparse Representation Classification (SRC). The researchers [3] purposed shape-based features extraction approach and for the classification task, random forest, and Support Vector Machine are used. In the literature, it was observed the performance of the ML classification methods heavily depends on the features extraction methods. Making the choice of appropriate features for supervised Learning tasks is a fatiguing and time-consuming process. Among many of the mentioned above, hand-crafted features are not robust. It can be observed that lots of hand-crafted features are not robust and usually their discriminative power is low [29–31]. Unlike these proposed methods our goal is to build a model based on Vision Transformer (ViT) [32] which is a very popular method in various areas of computer vision in recent times. The success of the Transformers [33] in Machine translation gives an idea of convolution-free models that basically depend on transformer layers, which makes this method interesting in the computer vision area. Moreover, Vision Transformer [32] models show potential to achieve higher accuracy equal to or even higher than convolutional neural networks-based models. Recently, there are various types of vision transformers-based methods are proposed such as pyramid structures like convolutional neural networks (CNNs) [34], data-efficient vision Transformer [35], and all-to-all self-attention [36]. In our proposed method, we used the vision transformer model as a feature extractor in order to take leverage of powerful machine learning classifiers. In the first step, Otsu’s Thresholding technique was utilized to remove noisy part of image. Then denoised training images are passed to the vision transformer for training and testing images features are extracted from this model. Where training of the machine learning classifier (decision tree) is done using extracted features from ViT model. The Fig. 1 indicates the flow of work to detect the multiclass classification. We input the multiclass (glioma, meningioma and pituitary) data, segmented the images partitioned the data with 70% for training and 30% for testing. The vision transformer is used for training and extracted features from vision transformer along with decision tree. The final model provides the multiclass classification results based on the vision transformer and vision transformer-decision tree (Fig. 2).
Improving Multi-class Brain Tumor Detection
5
Fig. 1. Schematic diagram
2 Materials and Methods 2.1 Dataset A publicly available dataset from (https://github.com/chengjun583/brainTumorRetri eval) was utilized in our study as described in [28, 37]. There were 233 patients from 3064 T1-weighted contrast enhanced images of three subtypes i.e. gliomas (1426 slices), pituitary tumor (930 slices) and meningiomas (708 slices). 2.2 Vision Transformer Transformer were proposed originally in [38]. The main objective to design transformer is to allow the modeling of dependencies irrespective to their distance in input sequence. The transformers due to the parallel computing capability are also more robust than the recurrent neural networks (RNNs) by removing recurrent connection. There are several encoder and decoder layers of transformer. An input to the high-level encoding is transformed by encoder whereas the decoder transforms an embedding to output. An encoder contained several encoding layers. For all encoded layers, the input is denoted as a tensor x to the encoder layer with a shape of T × C, here time steps are denoted by T and channels by C. Moreover, W Q denote the query transform matrix, W K represent the key transformed matrix and W V denote the value transform matrix. Likewise, C × dk shows the shape of W Q and W K and W V shape is C × dv , here dk and dv denote the integers. The query (K), key (K) and value (V) are obtained using: Q = xW Q
6
A. A. Abbasi et al.
Fig. 2. General Architecture of Vision Transformer
K = xW K V = xW V
(1)
Here T × dk shows shape of query Q and key K and value V has a shape T × dV . Thus, output of an encoder layer rewritten as: QK T V (2) h = softmax √ dk
Improving Multi-class Brain Tumor Detection
7
The output h has a shape of T ×H .√ The dot product of query with all keys is computed using this equation by dividing with dk . The softmax function is applied to obtained weights on value V. The dk square root is normalized term. Moreover, T × T denote the inner product of Q and K T represent feature correlation and different time steps. The correlation values are converted by SoftMax operation to probabilities along time steps indicating how much value V in a time step can be attended. The schematic diagram of the standard vision transformer is depicted in Fig. 1.
3 Results and Discussion In our study, the results of the vision Transformer and features extracted from the vision transformer model are presented in the below table. The performance of these methods was measured using standard measures (Figs. 3 and 5).
Fig. 3. Multi-class brain tumor type detection based on ResNet101-KNN
8
A. A. Abbasi et al.
Fig. 4. Multi-class brain tumor type detection based on Vision transformer
Figures above reflects the multi-class brain tumour classification using Residual neural network 101 (ResNet101)-KNN, Vision transformer, Vision transformer-DT. The deep learning ResNet101-KNN yielded the multiclass detection performance with accuracy (98.80%), sensitivity (98.29%), specificity (99.04%), PPV (97.96%), NPV (99.20%) and FPR (0.0095). The vision transformer yielded multiclass classification with accuracy (99.89%), sensitivity (99.65%), specificity (100%), PPV (100%), NPV (99.84%) and FPR (0.0000) to predict pituitary followed by glioma with accuracy (96.19%), sensitivity (99.07%), specificity (93.66%) and meningioma with accuracy (96.08%), sensitivity (84.42%), specificity (99.30%). The vision transformer with decision tree for feature extraction yielded highest detection performance to detect pituitary with accuracy (99.89%), sensitivity (100%), specificity (99.84%), PPV (99.62%), NPV (100%), FPR (0.0015) followed by glioma with accuracy (98.04%), sensitivity (99.56%), specificity (93.66%) and meningioma with accuracy (96.08%), sensitivity (84.42%), specificity (99.30%) (Figs. 6 and 7).
Improving Multi-class Brain Tumor Detection
9
Fig. 5. Multi-class brain tumor type detection based on Vision transformer-Decision tree (DT)
Fig. 6. Area under the receive operating curve (AUC) using Vision transformer
10
A. A. Abbasi et al.
Fig. 7. Area under the receive operating curve (AUC) using Vision transformer Decision tree (DT)
The Fig. 4 reflect AUC to distinguish the multiclass classification using a) Vision transformer, b) Vision transformer-DT. Using the vision transformer the highest separation was obtained to detect pituitary with AUC of 1.00 followed by glioma with AUC of 0.96 and meningioma with AUC of 0.92. Using the vision transformer-decision tree as feature extractor, the highest AUC was yielded to detect pituitary with AUC of 1.00 followed by glioma with AUC of 0.98 and meningioma with AUC of 0.96. Table 1. Results comparison utilizing different techniques on similar dataset Author
Feature/Methods
Performance
Afshar et al. [39]
CapsNets
Accuracy: 90.89%
Deepak and Ameer [40]
Deep GoogleNet-SVM
Accuracy: 97.10%
Swati et al. [41]
VGG-19 - Softmax
Sensitivity:96.81% Accuracy: 94.82% Specificity:93.93%
Anjum et al. [42]
CNN ResNet101 ResNet101-KNN
Sensitivity: 98.29% Specificity: 99.04% Accuracy: 98.80% AUC: 0.99
This study
Vision Transformer-Decision tree
Accuracy: 99.89% AUC: 1.00
Improving Multi-class Brain Tumor Detection
11
Multiclass brain tumor types of prediction is still a challenging task. Currently, researchers are developing different techniques to improve the multiclass classification. Table 1 summarizes the most recent methods developed to predict the multiclass brain tumor types using Fig share dataset. Recently Afser et al. [39] utilized CapsNet and obtained a highest accuracy of 90.89%. Swati et al. [41] utilized deep learning VGG-19 with accuracy of 94.82%. Deepak and Ameer [40] used deep learning GoogleNet-SVM and further improved accuracy of 97.10%. Anjum et al. [42] employed ResNet101-KNN and obtained an accuracy of 98.80% and AUC of 0.99. This study further improved the multi-class classification with accuracy of 99.89% and AUC of 1.00.
4 Conclusions Recently the researchers utilized different deep learning convolutional neural network (CNN) algorithms with default SoftMax and replaced softmax with SVM and KNN to improve the brain tumor multiclass classification. However, the present study, utilized a vision transformer with default and Decision tree for feature extraction. The result revealed that proposed approached further improved the multiclass classification. Thus the proposed approach is more robust than the previous methods utilized for multiclass brain tumor type detection. The results reveal that proposed method can be better utilized for improved prognosis, diagnosis and treatment planning. Abbreviations Magnetic resonance imaging (MRI) Scale-invariant Fourier transform (SIFT) Elliptic Fourier descriptors (EFDs) Machine Learning (ML) Space Pyramid Matching (SPM) K-nearest neighbor (KNN) Sparse Representation Classification (SRC) Vision Transformer (ViT) Convolutional neural networks (CNNs) Decision tree (DT) Recurrent neural networks (RNNs) Positive Predictive Value (PPV) Negative Predictive Value (NPV) Accuracy (ACC) Area Under the receiver operating characteristic Curve (AUC) Residual neural network 101 (ResNet101) False positive rate (FPR) Total accuracy (TA)
12
A. A. Abbasi et al.
References 1. Zhang, J., Shen, X., Zhuo, T., Zhou, H.: Brain tumor segmentation based on refined fully convolutional neural networks with a hierarchical dice loss (2017) 2. Sobhaninia, Z., Rezaei, S., Karimi, N., et al.: Brain tumor segmentation by cascaded deep neural networks using multiple image scales. In: 2020 28th Iranian Conference on Electrical Engineering, ICEE 2020 (2020). https://doi.org/10.1109/ICEE50131.2020.9260876 3. Asodekar, B., ProfS, G.: Brain tumor classification using shape analysis of MRI images. SSRN Electron. J. (2019). https://doi.org/10.2139/SSRN.3425335 4. Sobhaninia, Z., Rezaei, S., Noroozi, A., et al.: Brain tumor segmentation using deep learning by type specific sorting of images (2018) 5. Pashaei, A., Sajedi, H., Jazayeri, N.: Brain tumor classification via convolutional neural network and extreme learning machines. In: 2018 8th International Conference on Computer and Knowledge Engineering, ICCKE 2018, pp. 314–319 (2018). https://doi.org/10.1109/ ICCKE.2018.8566571 6. Gumaei, A., Hassan, M.M., Hassan, M.R., et al.: A hybrid feature extraction method with regularized extreme learning machine for brain tumor classification. IEEE Access 7, 36266– 36273 (2019). https://doi.org/10.1109/ACCESS.2019.2904145 7. Asim, Y., Raza, B., Malik, A.K., et al.: A multi-modal, multi-atlas-based approach for Alzheimer detection via machine learning. Int. J. Imaging Syst. Technol. 28, 113–123 (2018). https://doi.org/10.1002/IMA.22263 8. Rathore, S., Iftikhar, A., Ali, A., Hussain, M., Jalil, A.: Capture largest included circles: an approach for counting red blood cells. In: Chowdhry, B.S., Shaikh, F.K., Hussain, D.M.A., Uqaili, M.A. (eds.) IMTIC 2012. CCIS, vol. 281, pp. 373–384. Springer, Heidelberg (2012). https://doi.org/10.1007/978-3-642-28962-0_36 9. Rathore, S., Hussain, M., Khan, A.: Automated colon cancer detection using hybrid of novel geometric features and some traditional features. Comput. Biol. Med. 65, 279–296 (2015). https://doi.org/10.1016/J.COMPBIOMED.2015.03.004 10. Namekar, M., Ellis, E.M., O’Connell, M., et al.: Evaluation of a new commercially available immunoglobulin M capture enzyme-linked immunosorbent assay for diagnosis of dengue virus infection. J. Clin. Microbiol. 51, 3102–3106 (2013). https://doi.org/10.1128/JCM.003 51-13 11. Rathore, S., Hussain, M., Aksam Iftikhar, M., Jalil, A.: Ensemble classification of colon biopsy images based on information rich hybrid features. Comput. Biol. Med. 47, 76–92 (2014). https://doi.org/10.1016/J.COMPBIOMED.2013.12.010 12. Hussain, L., Ahmed, A., Saeed, S., et al.: Prostate cancer detection using machine learning techniques by employing combination of features extracting strategies. Cancer Biomark 21, 393–413 (2018). https://doi.org/10.3233/CBM-170643 13. Sundararaj, G.K., Balamurugan, V.: Robust classification of primary brain tumor in computer tomography images using K-NN and linear SVM. In: Proceedings of 2014 International Conference on Contemporary Computing and Informatics, IC3I 2014, pp. 1315–1319 (2014). https://doi.org/10.1109/IC3I.2014.7019693 14. Abdelaziz Ismael, S.A., Mohammed, A., Hefny, H.: An enhanced deep learning approach for brain cancer MRI images classification using residual networks. Artif. Intell. Med 102, 101779 (2020). https://doi.org/10.1016/J.ARTMED.2019.101779 15. Abd-Ellah, M.K., Awad, A.I., Khalaf, A.A.M., Hamed, H.F.A.: Design and implementation of a computer-aided diagnosis system for brain tumor classification. In: Proceedings of the International Conference on Microelectronics, ICM 0, pp. 73–76 (2016). https://doi.org/10. 1109/ICM.2016.7847911
Improving Multi-class Brain Tumor Detection
13
16. Sanjeev Kumar, P.M., Chatteijee, S.: Computer aided diagnostic for cancer detection using MRI images of brain (brain tumor detection and classification system). In: 2016 IEEE Annual India Conference, INDICON 2016 (2017). https://doi.org/10.1109/INDICON.2016.7838875 17. Islam, A., Hossain, M.F., Saha, C.: A new hybrid approach for brain tumor classification using BWT-KSVM. In: 4th International Conference on Advances in Electrical Engineering, ICAEE 2017, January 2018, pp. 241–246 (2017). https://doi.org/10.1109/ICAEE.2017.825 5360 18. Sachdeva, J., Kumar, V., Gupta, I., et al.: Multiclass brain tumor classification using GA-SVM. In: Proceedings - 4th International Conference on Developments in eSystems Engineering, DeSE 2011, pp. 182–187 (2011). https://doi.org/10.1109/DESE.2011.31 19. Mathew, A.R., Anto, P.B.: Tumor detection and classification of MRI brain image using wavelet transform and SVM. In: Proceedings of IEEE International Conference on Signal Processing and Communication, ICSPC 2017, January 2018, pp. 75–78 (2018). https://doi. org/10.1109/CSPC.2017.8305810 20. Menaka Devi, T., Ramani, G., Xavier Arockiaraj, S.: MR brain tumor classification and segmentation via wavelets. In: 2018 International Conference on Wireless Communications, Signal Processing and Networking, WiSPNET 2018 (2018). https://doi.org/10.1109/WISPNET. 2018.8538643 21. Deepa, S.N., Devi, B.A.: Neural networks and SMO based classification for brain tumor. In: Proceedings of the 2011 World Congress on Information and Communication Technologies, WICT 2011, pp. 1032–1037 (2011). https://doi.org/10.1109/WICT.2011.6141390 22. Sornam, M., Kavitha, M.S., Shalini, R.: Segmentation and classification of brain tumor using wavelet and Zernike based features on MRI. In: 2016 IEEE International Conference on Advances in Computer Applications, ICACA 2016, pp. 166–169 (2017). https://doi.org/10. 1109/ICACA.2016.7887944 23. Minz, A., Mahobiya, C.: MR image classification using adaboost for brain tumor type. In: Proceedings - 7th IEEE International Advanced Computing Conference, IACC 2017, pp. 701– 705 (2017). https://doi.org/10.1109/IACC.2017.0146 24. Chauhan, S., More, A., Uikey, R., et al.: Brain tumor detection and classification in MRI images using image and data mining. In: International Conference on Recent Innovations in Signal Processing and Embedded Systems, RISE 2017, January 2018, pp. 223–231 (2018). https://doi.org/10.1109/RISE.2017.8378158 25. Latif, G., Butt, M.M., Khan, A.H., et al.: Multiclass brain Glioma tumor classification using block-based 3D Wavelet features of MR images. In: 2017 4th International Conference on Electrical and Electronics Engineering, ICEEE 2017, pp. 333–337 (2017). https://doi.org/10. 1109/ICEEE2.2017.7935845 26. Deepa, A.R., Sam Emmanuel, W.R.: MRI brain tumor classification using cuckoo search support vector machines and particle swarm optimization based feature selection. In: Proceedings of the 2nd International Conference on Trends in Electronics and Informatics, ICOEI 2018, pp. 1213–1216 (2018). https://doi.org/10.1109/ICOEI.2018.8553697 27. Machhale, K., Nandpuru, H.B., Kapur, V., Kosta, L.: MRI brain cancer classification using hybrid classifier (SVM-KNN). In: 2015 International Conference on Industrial Instrumentation and Control, ICIC 2015, pp. 60–65 (2015). https://doi.org/10.1109/IIC.2015.715 0592 28. Cheng, J., Huang, W., Cao, S., et al.: Enhanced performance of brain tumor classification via tumor region augmentation and partition. PLoS ONE 10, e0140381 (2015). https://doi.org/ 10.1371/JOURNAL.PONE.0140381 29. Fehr, D., Veeraraghavan, H., Wibmer, A., et al.: Automatic classification of prostate cancer Gleason scores from multiparametric magnetic resonance images. Proc. Natl. Acad. Sci. 112, E6265–E6273 (2015). https://doi.org/10.1073/pnas.1505935112
14
A. A. Abbasi et al.
30. Abbasi, A.A., et al.: Detecting prostate cancer using deep learning convolution neural network with transfer learning approach. Cogn. Neurodyn. 14(4), 523–533 (2020). https://doi.org/10. 1007/s11571-020-09587-5 31. Cameron, A., Modhafar, A., Khalvati, F., et al.: Multiparametric MRI prostate cancer analysis via a hybrid morphological-textural model. In: Conference Proceedings - IEEE Engineering in Medicine and Biology Society 2014, pp. 3357–3360 (2014). https://doi.org/10.1109/EMBC. 2014.6944342 32. Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al.: An image is worth 16x16 words: transformers for image recognition at scale (2020) 33. Vaswani, A., et al.: Attention is all you need. In: Advances in Neural Information Processing Systems, vol. 30 (2017) 34. Wang, W., Xie, E., Li, X., et al.: Pyramid vision transformer: a versatile backbone for dense prediction without convolutions (2021) 35. Touvron, H., Cord, M., Douze, M., et al.: Training data-efficient image transformers & distillation through attention, pp. 10347–10357 (2021) 36. Wu, L., Liu, X., Liu, Q.: Centroid transformers: learning to abstract with attention (2021) 37. Cheng, J., Yang, W., Huang, M., et al.: Retrieval of brain tumors by adaptive spatial pooling and fisher vector representation. PLoS ONE 11, e0157112 (2016). https://doi.org/10.1371/ journal.pone.0157112 38. Kong, Q., Xu, Y., Wang, W., Plumbley, M.D.: Sound event detection of weakly labelled data with CNN-transformer and automatic threshold optimization. IEEE/ACM Trans. Audio Speech Lang. Process. 28, 2450–2460 (2020). https://doi.org/10.1109/TASLP.2020.3014737 39. Afshar, P., Plataniotis, K.N., Mohammadi, A.: Capsule networks for brain tumor classification based on MRI images and coarse tumor boundaries. In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1368–1372. IEEE (2019) 40. Deepak, S., Ameer, P.M.M.: Brain tumor classification using deep CNN features via transfer learning. Comput. Biol. Med. 111, 103345 (2019). https://doi.org/10.1016/j.compbiomed. 2019.103345 41. Swati, Z.N.K., Zhao, Q., Kabir, M., et al.: Brain tumor classification for MR images using transfer learning and fine-tuning. Comput. Med. Imaging Graph. 75, 34–46 (2019). https:// doi.org/10.1016/j.compmedimag.2019.05.001 42. Anjum, S., Hussain, L., Ali, M., et al.: Detecting brain tumors using deep learning convolutional neural network with transfer learning approach. Int. J. Imaging Syst. Technol. 32, 307–323 (2022). https://doi.org/10.1002/ima.22641
Measles Rash Disease Classification Based on Various CNN Classifiers Lohitha Rani Chintalapati, Trilok Sai Charan Tunuguntla, Yagnesh Challagundla(B) , Sachi Nandan Mohanty, and S. V. Sudha School of Computer Science and Engineering, VIT-AP University, Amaravati, Andhra Pradesh, India [email protected]
Abstract. One of the most thoroughly researched and well-documented nonlinear infectious disease dynamical systems is measles. Infants and young children are most likely to contract the immunizable disease measles. Measles is a highly commutable viral infection that has a 90% secondary infection incidence among contacts who are vulnerable. In this study, we have used a deep convolutional neural network to discriminate between various skin diseases and measles rash. The categorization performance of each individually optimized DL model across all of their ensembles has been presented using the specified dataset. We tested four optimizers, namely SGD, ADAM, RMSprop, and RAdam, on three considered models in order to further improve them. These models include VGG16, InceptionV3, and ResNeXt50, on which individual 10-fold cross-validation is done. The maximum average 10-fold cross-validation accuracy of 98.62%, 99.31% recall, and 99.32% F1 score were achieved by the optimised Inception V3 using the SGD optimizer. Finally, our predictive model offers a method for early detection to assist physicians in treating and enforcing new laws and regulations. Keywords: Convolutional Neural Network · Measles · InceptionV3 · SGD
1 Introduction Measles also called rubeola is a highly contagious respiratory infection caused by a virus named morbillivirus belonging to the paramyxovirus family [1]. Measles outbreaks can lead to epidemics that result in numerous fatalities. It is an airborne disease that readily transmits from one individual to another through many ways including coughs, sneezes, and direct contact with the mouth or nasal secretions of an infected person. Measles is a highly contagious disease where nine out of ten people who are not immune and are nearby an infected individual will be affected. Symptoms can be usually observed within 14 days of exposure to the virus. Once a person gets exposed to the virus then within the next 10 days to 14 days it usually results in a high fever, runny nose with red and watery eyes, and results in white spots developing inside the cheeks at the starting stage. Later once the virus develops and expands then slowly rashes erupt, firstly a red rash begins as flat red spots on the face which then spreads down the body, after © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2023 Published by Springer Nature Switzerland AG 2023. All Rights Reserved S. Nandan Mohanty et al. (Eds.): ICISML 2022, LNICST 470, pp. 15–27, 2023. https://doi.org/10.1007/978-3-031-35078-8_2
16
L. R. Chintalapati et al.
that small, popped bumps and raised white spots appear on top of flat red spots of the red rash. Rashes will only last for 5–7 days and eventually reduces. The rashes often appear 14 days after being exposed to the virus. General symptoms of measles include high fever, cough, runny nose, tiredness, inflamed eyes, white spots in the mouth, sore throat, sensitivity to light, muscle pain, and a red rash. Severe complications observed could include encephalitis, dehydration, blindness, severe diarrhea, severe respiratory infections like pneumonia and ear infections. Early and late-stage central nervous system (CNS) consequences following acute measles are significant and frequently deadly [11]. Due to their vulnerability to several illnesses and vitamin A deficiency, young children who are malnourished and have weak immune systems are more susceptible to developing severe measles. Children below the age of five and adults above the age of 30 have a higher risk of complications. Measles is most observed in Infants and children especially malnourished children and so is often believed to be the common age group affected by measles hence measles is often regarded as a childhood illness, even though measles is typically thought of as a childhood illness, it can infect anyone at any age [7]. Prior to the development of measles vaccines, 95%–98% of children by the age of 18 were infected with the virus [8]. A wide range of clinical signs, from a typical mild self-limiting infection to mortality, are seen in measles patients. Measles is still common in many underdeveloped countries, particularly in parts of Asia and Africa. Measles may be eradicated from a population; however, this needs coverage with two doses of vaccination at rates ranging from 93% to 95% of the population [9]. Nearly 95% of measles deaths occur in countries with poor healthcare infrastructure and low per capita income levels. Measles can infect anyone who hasn’t received a vaccination. Nearly everyone contracted the disease before the development of the measles vaccination. Measles, which was assumed to be a vanishing viral infection due to vaccination, has resurfaced globally. This has been linked to antivaccination campaigns in the early twenty-first century [2]. The measles vaccine is administered frequently in conjunction with other immunizations and has been proven to be safe and effective at avoiding the disease. Between 2000 and 2017, measles mortality decreased by 80% as a result of vaccination, and as of 2017, 85% of kids globally have received their first dose. Measles affects around 20 million people each year, primarily in underdeveloped countries in Africa and Asia. Measles killed 2.6 million people in 1980 and 545,000 in 1990; by 2014, global immunization initiatives had reduced measles deaths to 73,000. Despite these trends, sickness and death rates have rapidly increased from 2017 to 2019 due to a decline in immunization. Accurate virus diagnosis will be essential in reducing the number of illnesses as the virus spreads throughout the world. Because of how contagious measles is, outbreaks are a sign that the healthcare system needs help [3]. However, because the disease is less widespread than many others, it is more challenging to identify the virus, particularly for young medical professionals with minimal expertise. Since the other symptoms of measles are mostly interchangeable with those of other illnesses, the skin rash that it causes is its most distinctive characteristic. Healthcare professionals utilize the rash’s particular pattern and its distribution over the body to visually diagnose the illness. Across many disciplines, including medicine Artificial intelligence-based classification models have significantly changed how predictive decision-making is done [18]. Deep
Measles Rash Disease Classification
17
learning technologies, with their capacity to automatically learn semantic features from big datasets, are becoming useful in the detection of diseases [17]. In this study, we apply convolutional neural networks, a type of neural network model, to identify the measles rash using an image dataset by extracting higher representations of the visual information. We deploy a variety of CNN models, including VGG16, Inception V3, and ResNeXt50, along with optimizing techniques including SGD, ADAM, RMSprop, and Adam to help detect measles virus infection. The rest of the paper is structured as follows: Sect. 2 addresses related work, Sect. 3 talks about data, Sect. 4 discusses the methodology, Sect. 5 describes the results, and Sect. 6 concludes the paper.
2 Related Work Convolutional neural networks (CNNs) have achieved near or even superior performance than humans in the imaging sector [14]. Li, Ling-Fang, et al. findings show that deep learning-based skin disease image identification beats dermatologists and other computer-aided treatment approaches in skin disease diagnosis, with the multi-deep learning model fusion method having the maximum recognition effect [13]. Income level and Measles cases have a link. Geographic Information System (GIS) can help in disease prevention decision-making, such as in the case of measles [5]. Glock, Kimberly, et al. made use of transfer learning to construct Deep CNN to differentiate measles rash from various skin diseases. Analysis with the Residual Neural Network −50, which was trained on a varied and regulated array of skin rash images, produced 95.2% classification accuracy, 81.7% sensitivity, and 97.1% specificity. This shows that the strategy is effective in assisting with the detection of measles to aid in the containment of outbreaks [6]. Wu, Z. H. E., et al. discovered that transfer learning models have higher average precision and recall [14]. Several deep CNN architectures are proposed to investigate the potential of Deep Learning trained on the “DermNet” dataset for the diagnosis of 23 different types of skin diseases. These architectures are compared to determine which one performs the best. The method demonstrates that DenseNet was the most accurate for skin disease classification using the DermNet Dataset, with a Top-1 accuracy of 68.97% and a Top-5 accuracy of 89.05% [10]. To diagnose three skin illnesses, the classification performances of five deep network architectures were investigated. Comparisons were made using the accuracy, specificity, precision, F1 metric, and MCC. According to quantitative results, ResNet101 can classify images more accurately than the other networks [15]. According to the experimental findings of the classification of Pneumonia using Inception-V3 and Convolutional Neural Networks from X-Ray Images by Mujahid, Muhammad, et al., Inception-V3 with CNN achieved the highest accuracy and recall scores, at 99.29% and 99.73%, respectively [16]. Ali, Shams Nafisa, and colleagues demonstrated how to extract picture features using a convolution neural network and the deep learning idea. To classify monkeypox and other diseases, multiple pre-trained deep-learning models, including VGG-16, ResNet50, and InceptionV3 are used. A combination of the three models is also created. ResNet50 achieves the highest overall accuracy of 82.96 (4.57%), while VGG16 and the ensemble system achieve 81.48 (6.77%) and 79.26 (1.05%), respectively. As an online monkeypox screening tool, a prototype web application is currently being created [12].
18
L. R. Chintalapati et al.
There is a need to precisely identify measles rash from the images of the various skin diseases and provide the required treatment on time to reduce the complications and death rate. Improving the ability to identify measles would benefit healthcare professionals in tackling the impending measles problem and the prospect of measles returning to the country, the dermatologists are in scarce supply in many areas and hence deep learning and image classification concepts can improve public health in many poor nations. Deep learning-based convolutional neural network models prove to have a good potential in detecting various skin diseases and have been effectively used by various researchers; improvising on the above, we propose to use three different convolutional neural network models, namely VGG-16, ResNeXt-50, and Inception V3, with four different optimizers, namely SGD, ADAM, RMSprop, and Adam, on which we implement a 10-fold crossvalidation for better predictions.
3 Data An image dataset of various skin conditions and rashes was downloaded from IEEE-Data Port. The dataset contains images of rashes of 11 different diseases. The dataset also contains pictures of people without any skin infections. There are two folders called “Train” and “Test” under the directory “images”. In the “Train” folder, there are two sub-folders named “Measles” and “Not Measles”, each containing 126 and 926 images respectively. In the “Test” folder, there are two sub-folders named “Measles” and “Not Measles”, each containing 32 and 232 images respectively. It has images containing 11 different rashes and normal skin. A detailed description of the dataset is represented in the table below Table 1. Table 1. Information about our dataset Image Class
# of images
Chickenpox
170
Measles
158
Ringworm
131
Bowen’s Disease
124
Psoriasis
122
Enterovirus
117
Keratosis
112
Eczema
95
Chigger Bites
87
Dermatofibroma
80
Scabies
79
Normal Skin
41
Total
1316
Measles Rash Disease Classification
19
The following pictures displays two illustrations of our dataset, Fig. 1 shows Measles rash and Fig. 2 shows Chickenpox rash. The similarity in appearance of the two types of rashes represents a challenge in differentiating the measles rash from other skin problems.
Fig. 1. Measles Rash
Fig. 2. ChickenPox Rash
4 Methodology 4.1 Model Convolutional Neural Network is an artificial deep learning neural network model applied to analyse visual imagery. The CNN architecture has demonstrated outstanding performance in a wide range of Computer Vision and Machine Learning challenges [4]. Different layers in CNN are the Convolution layer, Pooling layer, ReLU layer, and fully connected layer. In this study, we have implemented 3 different types of CNN Architectures namely VGG16, Inception V3, and ResneXt50, and applied SGD, ADAM, RMSprop, and Adam optimizers which are used to reduce the overall loss and improve the accuracy. 4.1.1 VGG-16 The VGG model stands for the Visual Geometry Group from Oxford. An illustration of VGG-16 model is shown in Fig. 3. This model is far more in-depth than AlexNet and it is quite intuitive. A convolutional neural network with 16 deep layers is called VGG-16. Its image input dimensions are 224 by 224. In the considered dataset there are two distinct folders—train and test folders—have been created for the pictures. The train and test folders are split so that 80% of the data’s photos are utilized for training and 20% are used for testing. To execute 10-fold cross-validation, all libraries were installed from Keras. The input size and weights of the image with type = 32 were used to build the VGG model. Therefore, a package called glob is used to mount the folders from the training sample from the drive to train the model. The model includes three layers. In Layer 1, there are two CONV 2D layers and a MaxPooling layer. In Layer 2, there are two CONV 2D layers and a Pooling layer, and in Layer 3, there are three CONV 2D layers and a Max, with the other two layers having the same format and being followed by a dense layer.
20
L. R. Chintalapati et al.
Fig. 3. Illustration Of VGG-16 Model
There are 14,764,866 parameters altogether. 50,178 are the trainable parameters, and 14,714,688 are the non-trainable parameters. After the model has been run, the model.h5 file is created automatically. Next, using the Keras Image Data Generator with the parameter’s width shift range argument = 0.2, rotation range argument = 40 and height shift range argument = 0.2, train and test sets have been created. In the train set, we discovered 1050 images that belonged to two classes, while in the test set, we discovered 264 images that belonged to two classes. Model.h5 has now been utilized to assess the model and determine all potential metrics by building a confusion matrix. 4.1.2 ResNeXt50 ResNeXt replaces the residual block with a “split-transform-merge” strategy, which is similar to Inception’s module but differs in the aspect that Inception employs a separate filter and size for each block, ResNeXt uses shared hyper-parameters. It also uses a much more parallel stacking layer than sequential layers. ResNeXt is named the ILSVRC 2016 classification task’s first Runner Up. The learning ratio is low when compared to AlexNet and VGG because those architectures do not provide batch normalization. This model corresponds to a single node with 8 GPUs, learning ratio of 0.1 and batch size = 32. The libraries are used to assess the model image classification. To train the model, batch normalization is performed on the input, then a 1x1 convolution layer is added, followed by max pooling, then a four-stage layer of grouped convolution, global average pooling, residual connection, and max pooling to create a fully connected layer. On the whole, in_features are 2048 and out_features are 1000 where bias is True. An illustration of ResNeXt50 model is shown in Fig. 4. 4.1.3 Inception V3 For image classification, Inception V3 is performed which is a simplified version of Google Neural Network where it is employed. This type employs several filters of various
Measles Rash Disease Classification
21
Fig. 4. Figure Of The ResNeXt50 Model
sizes. Inception V3 is a better model adaptation since it utilizes regularizers with higher efficiency and uncompromised speed. Among other advancements, the Inception-v3 convolutional neural network architecture makes use of Label Smoothing, factorized 7 × 7 convolutions, and the addition of an auxiliary classifier to move label information lower down the network. The Inception V3 model has 42 layers in total, which is slightly more than the inception V1 and V2 models. Traditionally, the grid size of the feature maps was decreased using maximum and average pooling. The activation dimension of the network filters is enhanced in the Inception V3 model to decrease the grid size more. In comparison to its competitors, Inception V3 may attain the lowest error rates.
Fig. 5. Inception V3 architecture
By examining Fig. 5, we can comprehend the Inception V3 architecture. After Concatenate in the Inception V3 model, it flows through Dense, Dropout, and ultimately Dense Soft Max Layer.
22
L. R. Chintalapati et al.
4.2 Optimizers In Deep Learning Neural Networks, optimizers are used to change model’s parameters. An optimizer’s job is to reduce model weights to minimize a loss function. The loss function is used to analyse the model’s performance. A neural network model must be trained using an optimizer. The algorithm’s foundation is randomness, which is mentioned as stochastic. Rather than using the entire dataset for each iteration in stochastic gradient descent (SGD), we elect the data batches at random. This suggests that we only sample a small portion of the dataset. Adam may be a different optimization algorithm that can be used to train deep learning models instead of stochastic gradient descent. Adam creates an optimization technique which will handle sparse gradients in noisy situations by combining the best features of the AdaGrad and RMSProp algorithms. Root Mean Squared Propagation, or RMSProp, may be a variation on gradient descent, and therefore the AdaGrad version of gradient descent adapts the step size for each parameter using a declining average of partial gradients. Rectified Adam, often known as Adam, is a stochastic optimizer variation that adds a term to correct the adaptive learning rate’s variance. It attempts to solve Adam’s terrible convergence issue. The three CNN models considered have been trained using four optimizers (i.e., ADAM, RSMprop, SGD, RAdam). The higher accuracy of the 10-fold cross-validation and testing data is served as basis for separation between these 12 variations. A few further tests were performed on the most accurate model to evaluate how it performed under various conditions.
5 Results and Discussion We may infer from Table 2 that, among all the models with optimizers applied to them, the Inception V3 model with the SGD optimizer has the best average 10-fold crossvalidation Accuracy (98.62%). The average 10-fold cross-validation accuracy is less than 90%, for ResNeXt50 with SGD optimizer. VGG16 model with Adam optimizer achieves the second highest average 10-fold cross-validation accuracy. The RAdam optimizer achieves the highest average 10-fold cross-validation accuracy of 97.9% in the ResNeXt50 model.
Measles Rash Disease Classification
23
Table 2. Results using four optimizers for VGG16, InceptionV3 and ResNeXt50. Model
Optimizer
Average 10-Fold Cross Validation Accuracy (%)
VGG16
SGD
96.91
ADAM
98.59
RMSprop
97.40
Inception V3
ResNeXt50
RAdam
98.50
SGD
98.62
ADAM
96.95
RMSprop
97.77
RAdam
96.43
SGD
89.35
ADAM
97.48
RMSprop
97.68
RAdam
97.90
Precision is a metric used to analyse a model’s dependability and its accuracy in categorizing a sample as positive. It is calculated using the ratio of True Positives to True Positives and False Positives. Precision =
Truepositive Truepositive + Falsepositive
(1)
The Recall parameter is used to assess how well the model can identify positive test data. It is determined by the ratio of Positive samples that were correctly labelled as positive to the total number of positive Instances samples. The greater the Recall value, the greater the number of positive samples detected. Recall =
Truepositive Truepositive + FalseNegative
(2)
Models for classification include the F1 Score. The F1 Score is focused on precision and recall. The Harmonic mean of Precision and Recall is the F1 Score. F1Score = 2 ∗
Precision ∗ Recall Precision ∗ Recall
(3)
24
L. R. Chintalapati et al. Table 3. Precision, Recall, F1-score of the considered models with optimizers
Model
Optimizer
Precision (%)
Recall (%)
F1-score (%)
VGG16
SGD
97.92
98.08
97.99
ADAM
99.34
99.31
99.32
RMSprop
98.62
99.67
99.14
RAdam
98.8
99.34
99.07
SGD
99.34
99.31
99.32
ADAM
96.83
96.44
96.63
RMSprop
99.54
96.53
95.01
RAdam
99.01
98.71
98.25
SGD
96.49
96.39
96.43
ADAM
99.18
99.15
99.16
RMSprop
99.34
99.31
99.32
RAdam
99.14
99.08
99.11
Inception V3
ResNeXt50
From Table 3 we can observe the highest Precision of 99.54% was obtained for Inception V3 with RMSprop as optimizer. The highest Recall of 99.67% has been observed for VGG16 with RMSprop as optimizer. VGG16 with ADAM optimizer, Inception V3 with SGD optimizer, and ResNeXt50 with RMSprop optimizer have the highest F1 Scores, each recording 99.32%.
Fig. 6. F1-Score Recall Curve
Fig. 7. Accuracy Values Versus Epochs
We have plotted Recall against F1-Score in Fig. 6. The Validation Accuracy increases as the number of Epochs increases as observed in Fig. 7. The most popular cost function is called Cross Entropy Loss. It is used to reduce the loss. The output of cross-entropy loss, also known as log loss, which produces probability value between 0 and 1, is used to assess the effectiveness of classification models. −(ylog(p) + (1 − y)log(1 − p))
(4)
Measles Rash Disease Classification
25
Validation loss
1
Loss
0.8 0.6 0.4 0.2 0 0
5
10
20
25
30
Epochs VGG16
InceponV3
ResNeXt50
Fig. 8. Loss function values plotted against the Number of Epochs
As validation loss increases, we usually observe overfitting. In order to reduce overfitting, we use cross-entropy loss function Fig. 8. Depicts that as the number of Epoch increases then validation initially decreases, at a particular point it remains constant and then decreases again.
Fig. 9. Plots on application of various optimizers on considered CNN models
From Fig. 9 we can see that by utilizing any of the following optimizers SGD, RMSprop, ADAM and RAdam; InceptionV3 has obtained the highest average 10-fold cross-validation accuracy.
26
L. R. Chintalapati et al.
6 Conclusions Most measles cases range from mild to quite severe. But measles can also cause encephalitis, pneumonia, and even fatalities. Because the measles rash gradually spreads to the hands and feet, a skilled model is needed to identify the rash and take the appropriate safety measures to avoid fatality. We conducted research utilizing three different deep neural network architectures (VGG16, InceptionV3, and ResNeXt50) and four different optimizers (ADAM, RAdam, RMSprop, and SGD) to distinguish between images of the measles and those that were not. Inception V3 with SGD optimizer displayed the highest average 10-fold cross-validation accuracy of 98.62% with 99.31% Recall and 99.32% F1 Score among the 12 model versions we looked at as well as the results are listed in Tables 2 and 3. In the future, we will use autoencoders because Convolutional Neural Networks needs a huge sample data for training the model, but we only have a restricted number of photos. We learn about low-dimensional data representations with their help by selecting relevant characteristics with little adjustment and relying on confidence intervals rather than point estimations. For our CNN model to be more effective in diagnosing human skin diseases, it could also be required to expand our dataset to detect a wider range of rash disorders. We intend to investigate further optimizers in the future to enhance system performance even more. Additionally, we want to create a web application that will let users submit photos captured with their cameras or from libraries of images. The Inception V3 model examines supplied images to determine whether they are measlespositive or negative. If positive, a message such as “MEASLES DETECTED, PLEASE SEEK MEDICAL ATTENTION IMMEDIATELY” is displayed along with the model accuracy. By creating a web application which can be utilized as a dynamic tool to treat measles disease as the frequency of the disease grows, doctors can take advantage of the advancement of technology.
References 1. Bellini, W.J., Rota, J.S., Rota, P.A.: Virology of measles virus. J. Infect. Dis. 170(Suppl._1), S15–S23 (1994) 2. Xerri, T., et al.: Complications of measles: a case series. BMJ Case Rep. CP 13(2), e232408 (2020) 3. Takahashi, S., et al.: Reduced vaccination and the risk of measles and other childhood infections post-Ebola. Science 347(6227), 1240–1242 (2015) 4. Hussain, M., Bird, J.J., Faria, D.R.: A study on CNN transfer learning for image classification. In: Lotfi, A., Bouchachia, H., Gegov, A., Langensiepen, C., McGinnity, M. (eds.) UKCI 2018. AISC, vol. 840, pp. 191–202. Springer, Cham (2019). https://doi.org/10.1007/978-3319-97982-3_16 5. Sulistyawati, S., Sumiana, S.: Measles cluster detection using ordinal scan statistic model. Mater. Socio-Med. 30(4), 282 (2018) 6. Glock, K., et al.: Measles rash identification using transfer learning and deep convolutional neural networks. In: 2021 IEEE International Conference on Big Data (Big Data). IEEE (2021) 7. Sabella, C.: Measles: not just a childhood rash. Clevel. Clin. J. Med. 77(3), 207–213 (2010)
Measles Rash Disease Classification
27
8. Perry, R.T., Halsey, N.A.: The clinical significance of measles: a review. J. Infect. Dis. 189(Suppl._1), S4–S16 (2004) 9. Bester, J.C.: Measles and measles vaccination: a review. JAMA Pediatr. 170(12), 1209–1215 (2016) 10. Aboulmira, A., Hrimech, H., Lachgar, M.: Comparative study of multiple CNN models for classification of 23 skin diseases. Int. J. Online Biomed. Eng. 18(11) (2022) 11. Schneider-Schaulies, J., ter Meulen, V., Schneider-Schaulies, S.: Measles infection of the central nervous system. J. Neurovirol. 9(2), 247–252 (2003) 12. Ali, S.N., et al.: Monkeypox skin lesion detection using deep learning models: a feasibility study. arXiv preprint arXiv:2207.03342 (2022) 13. Li, L.-F., et al.: Deep learning in skin disease image recognition: a review. IEEE Access 8, 208264–208280 (2020) 14. Wu, Z.H.E., et al.: Studies on different CNN algorithms for face skin disease classification based on clinical images. IEEE Access 7, 66505–66511 (2019) 15. Goceri, E., Karakas, A.A.: Comparative evaluations of CNN based networks for skin lesion classification. In: 14th International Conference on Computer Graphics. Visualization, Computer Vision and Image Processing (CGVCVIP), Zagreb, Croatia (2020) 16. Mujahid, M., et al.: Pneumonia classification from X-ray images with inception-V3 and convolutional neural network. Diagnostics 12(5), 1280 (2022) 17. Sahu, B., Mohanty, S.N.: CMBA-SVM: a clinical approach for Parkinson disease diagnosis. Int. J. Inf. Technol. 13(3), 647–655 (2021). https://doi.org/10.1007/s41870-020-00569-8. ISSN: 2511-2104 18. Sahu, B.P., Mohanty, S.N., Rout, S.K.: A hybrid approach for breast cancer classification and diagnosis. EAI Endorsed Trans. Scalable Inf. Syst. 6(20), 1–8 (2019)
Brain Imaging Tool in Patients with Trans Ischemic Attack: A Comparative Research Study Analysis of Computed Tomography and Magnetic Resonance Imaging R. Bhuvana(B)
and R. J. Hemalatha
Department of Biomedical Engineering, VISTAS, Pallavaram, Chennai 6000117, Tamil Nadu, India [email protected], [email protected]
Abstract. The processing and analysis of brain imaging to identify transient ischemic strokes has remained difficult due to the requirement for more precise abnormality identification and the extraction of concealed but essential information from image data. This is necessary in order to diagnose transient ischemic strokes. Because of both of these conditions, identifying people who have had transient ischemic strokes has become more challenging. In order to arrive at a diagnosis of transient ischemic stroke, it is necessary to have fulfilled either one of these conditions. The work that is being done right now has the intention of achieving a higher level of precision in the process of extracting and selecting features from image data. The work that is being done right now places a significant emphasis on this particular aspect. This is being done in order to obtain a more in-depth understanding of the images in terms of the detection of abnormalities, and it is being done so right now. By analysing multiple groups of abnormalities side by side, the purpose of this research is to help advance the development of MRI and CT scans that are more accurate. The comparison of several different types of anomalies is the primary focus of this research. Keywords: Magnetic Resonance Imaging (MRI) · Computed Tomography (CT) · Transient Ischaemic Stroke (TIS) · Haemorrhage · Lesion
1 Introduction With millions of deaths reported each year, cerebrovascular illness or cerebrovascular disease is a major cause of morbidity as well as a leading cause of mortality. Three months after the stroke, 20% of stroke patients still need institutional care, and up to 50% of stroke survivors never achieve functional independence [1]. The development of neurological side effects following a cerebrovascular stroke is unavoidable and recommends investigation of an ischemic sore in the spinal cord or other non-ischemic etiology. The beginning could come on unexpectedly or after excitement, © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2023 Published by Springer Nature Switzerland AG 2023. All Rights Reserved S. Nandan Mohanty et al. (Eds.): ICISML 2022, LNICST 470, pp. 28–34, 2023. https://doi.org/10.1007/978-3-031-35078-8_3
Brain Imaging Tool in Patients with Trans Ischemic Attack
29
as well as cerebral pain, sickness, wooziness, or loss of cognizance. The two primary kinds of stroke are ischemic and hemorrhagic [4]. Treatment techniques for these two subtypes of stroke are particularly unique, and the early conclusion of stroke as well as assurance of the subtype is a significant early move toward stroke treatment [6, 7]. Many of the stroke cases are hemorrhagic strokes, otherwise called intracerebral hemorrhages. Not withstanding having a low frequency, this sort of stroke is related to significant morbidity and mortality. In the 30 days following a hemorrhagic stroke, up to 38% of patients will die [2, 3], and around half of survivors will in any case need help with regular exercise. Ischemic stroke is more normal, addressing roughly 85% of all stroke cases, and has a 30-day death rat, that is significantly lower, at around 12% [9]. Ischemic stroke’se horribleness can likewise be significant, and early finding and treatment are essential for limiting it. A hemorrhagic difference in an ischemic sore is likewise conceivable. This is particularly common for venous impediment related dead tissue. At the point when hemorrhagic dead tissue is found, venous (sinus) framework blockage ought to be considered. A transient ischemia assault (TIA), which causes a central neurological disability that recovers in 24 h, may likewise occur. Albeit self-restricted, TIA could make an ischemic stroke more troublesome. Up to 20% of TIA patients may encounter a stroke within a span of 90 days, making TIA, a critical momentary risk factor for stroke [12].
2 Diagnosis Methods 2.1 The Imaging Modality of Stroke Neuroimaging is helpful for transient ischemic stroke clinical finding, treatment, and treatment as well as anticipation evaluation. Years and years prior, the foundations of indicative imaging were radioisotope, thermography, and electroencephalography (EEG) techniques. In any case, the formation of CT and MRI, which capture pictures of the human mind, denoted a leap forward in imaging [15, 18]. These neuroimaging procedures’ essential goal is to distinguish the infarct centre and obscuration as well as the distressed vascular locale of a stroke patient’s cerebrum, as doing so would assist with decreasing the seriousness of the stroke by utilising the best treatment [25]. The clinical determination, diagnosis, and treatment of transient ischemic stroke as well as the assessment of stroke survivors rely heavily on neuroimaging. For this reason, the principal objective of these neuroimaging procedures is to pinpoint the impacted vascular district of a stroke patient’s cerebrum, the infarct centre, and the infarct obscuration [22]. Functional and structural neuroimaging techniques are the two categories. The purpose of structural imaging is to visualise the various anatomical structures of the brain and any deformities related to them, such as a tumour, clot, or bleeding, whereas the purpose of functional imaging is to assess activity in various regions of the brain. Functional neuroimaging comprises functional MRI (fMRI), PET, and other functional neuroimaging techniques, whereas structural neuroimaging includes CT and MRI [5–7].
30
R. Bhuvana and R. J. Hemalatha
3 Computed Tomography CT is a neuroimaging technology that is widely used because it is relatively affordable, has a faster imaging time, and is more accessible than MRI [8]. According to a study, a CT scan employs detectors and X-rays to produce cross-sectional images of the brain. The contrast of CT images is impacted by the differing ways that various tissue types absorb X-rays. CT scans provide high-resolution images of osseous or bony structures rather than soft tissue. It is therefore frequently suggested for bone fracture imaging [20]. Various CT modalities are currently frequently utilised in the conclusion of TIAs because of the continuous advancements in CT innovation. Numerous techniques, including head CT, are utilized, relying upon the patient’s resilience to differentiate specialists and assess the hazards of anticipation. [13] Numerous procedures, including head CT, computed tomography perfusion (CTP), processed tomography angiography (CTA), or joint application, are vigorously utilised to get the exhaustive data of the patient at a time, depending on the patient’s resilience to different specialists and risk evaluation of the forecast [24]. CT angiography (CTA) is a method for showing the state of a vein, like the state of a vein and a stenosis, by handling a picture on a CT scan procedure when the difference specialist develops at the vein imperfection [14, 15]. The utilisation of CTA imaging in deciding the anticipation of strokes has been the subject of various examinations. People with impediment or high-grade stenosis on the brain who undergo CTA present with prompt stroke side effects and have an unfortunate forecast [16]. In patients with intense stroke side effects, a head CT scan cannot recognise stroke or TIA. As of late, CTA has shown itself to be a helpful instrument for deciding if there is significant stenosis or impediment [20]. CTP uses dynamic results with contrast experts to detect intracerebral hemodynamics at explicit levels [18]. The advantages of plain CT really look at the ability of grant CTP to examine back-course TIAs even more unequivocally, which is maintained by the assurance stream diagnostic test [19]. Since CTP provides different benefits, it also has some basic drawbacks, such as the low explicitness of the evaluation, enormous assortments, and limits in the perfusion limits [23]. A couple of examinations uncovered that CTA and CTP coupled can construct the accuracy of predicting promises [17]. Future insightful examinations should be coordinated, yet getting the data for them can be challenging. The table below provides insight on the various CT modalities used in the study of transient ischemic stroke (Table 1).
4 Magnetic Resonance Imaging Significant standard images of the body’s fragile tissues, including the frontal cortex, are made by MRI, a multimodal imaging gadget used to investigate life frameworks and capacities [10]. Also, diverging from CT, it offers transcendent tissue contrast. MRI takes advantage of the tissues’ hydrogen centers, which have alluring properties, to picture actual regions’ inside plans. Different MRI sequences are used in neuroimaging to visualise explicit brain regions [11].
Brain Imaging Tool in Patients with Trans Ischemic Attack
31
Table 1. The various CT modalities used in the study of transient ischemic stroke MODALITY TYPE
STRENGTH
LIMITATIONS
REFERENCE
NCCT
non-invasive, lower cost
Radiation exposure is less efficient in diagnosis
Menon et al. 2015
CTA
Invasive
Contrast agents can harm the kidney
Menon et al. 2015
CTP
Faster than perfusion MRI
Less anatomical analysis, allergy due to contrast agents
Wing and Mark 2019; Zhang et al. 2016
Transient ischemic stroke (TIS) can also be identified using magnetic resonance imaging (MRI) [17]. Additionally, MRI includes a variety of patterns to achieve the best assessment for individuals who are classified according to their unique risk profile [11]. Studies have indicated that despite the benefit of its broad applicability, it overestimates the intracranial pressure (ICP) or changes in cerebral blood flow (CBF). The clinical procedure known as perfusion-weighted imaging (PWI) is most commonly utilised and has a serious level of precision [26]. As per the imaging guideline of PWI, which expects to address changes in cerebrum microvascular morphology and micro hemodynamics of the brain, the nearby mean travel time (MTT) is equivalent to the neighbourhood cerebral blood volume (CBV) [27]. Regarding CVR, PWI, and CTP, they are not noticeably different from one another. On parallel and unaffected brain tissue, the rMTT can measure changes in local blood flow and identify early ischemic tissue. A past report tracked down that new cerebrum localised necrosis (BI), which is pertinent to repetitive cerebral dead tissue, is connected to 30% of intense central PWI injuries after TIA. Diffusion-weighted imaging (DWI), which most importantly measures the diffusion of water molecules, can quickly identify acute vascular changes after an obstruction. The biomarker, the Diffusion Coefficient (ADC) created by the DWI sequence, has shown to be able to distinguish between normal, ischemia-free, and obstructed areas in the brain. By comparing their ADC readings, radiologists can discriminate between infarcted brain tissue and penumbra regions, which may provide all the necessary information to assist the diagnosis and treatment techniques for certain patients. Additionally, one researcher observed that DWI has a higher accuracy rate than CT in his research on the analysis of computed tomography and magnetic resonance imaging. Additionally, the American Stroke Association Committee advises using DWI as a first choice for those who have TIA symptoms due to its great sensitivity. The table below provides insight on the various MRI modalities used in the study of transient ischemic stroke (Table 2). .
32
R. Bhuvana and R. J. Hemalatha
Table 2. Insights on the various MRI modalities used in the study of transient ischemic stroke MODALITY TYPE
STRENGTH
LIMITATION
REFERENCE
ACCURACY
T1 and T2 weighted images
Clear anatomical analysis
Vasculature is unclear
Saad et al., 2015
90%
DWI
very high sensitivity
False negative results
Okariec et al., 2015
92%
PWI
very good sensitivity
Contrast creates allergies
Demeestree et al., 2020
97%
4.1 Anatomy of CT and MRI in Stroke Imaging According to a study, multimodal magnetic resonance technology yields outcomes in the response to acute haemorrhage that are comparable to those of a CT scan [15]. Furthermore, haemorrhages that show up as acute on a CT scan could show up as chronic on an MRI, making a CT scan a good choice for making the first diagnosis of a patient with an acute stroke [17]. Additionally, some investigations have shown that the MRI method is superior to the CT in identifying cerebral microbleeds and persistent haemorrhages. However, because of its extended scanning period, MRI may be susceptible to artefacts brought on by body movements [18]. 4.2 CT Used to Image a Brain Stroke Quick information obtaining, straightforward entry, understanding, the capacity to picture fundamentally sick, claustrophobic, or upset patients, as well as amazing responsiveness for distinguishing intracranial drain, are benefits of CT neuroimaging over X-ray [24]. Nonetheless, CT likewise has impediments, including the gamble of openness to ionizing radiation and hypersensitive responses to differentiate specialists, particularly in patients with diabetes and renal sickness [21]. 4.3 MRI is Used to Image a Brain Stroke While MRI has an advantage over CT in regards to perceiving cerebral ischemic stroke, this includes higher costs, longer result times, and a riotous, tunnel-like arrangement that can make it difficult to really take a look at patients with essential illnesses, disquiet, or claustrophobia, as well as individuals who have inbuilt pacemakers, aneurysm bruises, or other metallic things implanted in their bodies [15]. Nevertheless, the physiological information amassed from MRI can altogether assist in the clinical administration of aid and the clinical organisation of stroke patients by definitively revealing the size of the infarct with coring, the obscuration, and the site of hindrance [26]. To affirm the usefulness of MRI for coordinating careful and exact treatment, by colossal extension, clinical assessments are currently required [24].
Brain Imaging Tool in Patients with Trans Ischemic Attack
33
5 Conclusion With the guide of picture information examination and perception, the most generally involved clinical imaging modalities for the identification of neurological issues related to transient ischemic stroke and the conclusion of dying, ischemic stroke, and malignancies are presently magnetic reverberation imaging (MRI) and computed tomography (CT). Unrivaled delicate tissue separation, difference, and incredible spatial goals are totally upheld by MRI, which doesn’t produce risky ionising radiation [17]. These qualities have made MRI a significant apparatus for finding, especially in clinical and careful settings, early strokes (transient ischemic stroke) and malignancies [24]. Another methodology, CT filters, is essentially used to recognise ischemic stroke, which is brought on by a limitation of blood flow to the cerebrum, and draining stroke, which is brought on by draining in the brain. The quicker examining span of a CT check works with delicate data, making it better for the early conclusion of draining and stroke.
References 1. Lloyd-Jones, D., et al.: Heart disease and stroke statistics-2009 update. Circulation 119(3), e21–e181 (2009) 2. Rosamond, W.D., et al.: Stroke incidence and survival among middle-aged adults: 9-year follow-up of the atherosclerosis risk in communities (ARIC) cohort. Stroke 30(4), 736–743 (1999) 3. Rost, N.S., et al.: Prediction of functional outcome in patients with primary intracerebral hemorrhage. Stroke 39(8), 2304–2309 (2008) 4. Astrup, J., Siesjö, B.K., Symon, L.: Thresholds in cerebral ischemia - the ischemic penumbra. Stroke 12(6), 723–725 (1981) 5. Heiss, W.D.: The ischemic penumbra: correlates in imaging and implications for treatment of ischemic stroke. Cerebrovasc. Dis 32(4), 307–320 (2011). The Johann Jacob Wepfer award 2011 6. Hossmann, K.A.: Periinfarct depolarizations. Cerebrovasc. Brain Metab. Rev. 8(3), 195–208 (1996) 7. Dohmen, C., et al.: Spreading depolarizations occur in human ischemic stroke with high incidence. Ann. Neurol. 63(6), 720–728 (2008) 8. Wildermuth, S., Knauth, M., Brandt, T., Winter, R., Sartor, K., Hacke, W.: Role of CT angiography in patient selection for thrombolytic therapy in acute hemispheric stroke. Stroke 29(5), 935–938 (1998) 9. Barber, P.A., et al.: Prediction of stroke outcome with echoplanar perfusion - and diffusionweighted MRI. Neurology 51(2), 418–426 (1998) 10. Merwick, A., et al.: Addition of brain and carotid imaging to the ABCD 2 score to identify patients at early risk of stroke after transient ischaemic attack: a multicentre observational study. Lancet Neurol. 9, 1060–1069 (2010) 11. Moreau, F., et al.: Early magnetic resonance imaging in transient ischemic attack and minor stroke: do it or lose it. Stroke 44, 671–674 (2013) 12. Coutts, S.B., Modi, J., Patel, S.K., Demchuk, A.M., Goyal, M., Hill, M.D.: CT/CT angiography and MRI findings predict recurrent stroke after transient ischemic attack and minor stroke: results of the prospective catch study. Stroke 43, 1013–1017 (2012) 13. Steffenhagen, N., et al.: Reliability of measuring lesion volumes in transient ischemic attack and minor stroke. Stroke 41, 814–816 (2010)
34
R. Bhuvana and R. J. Hemalatha
14. Lev, M.H., et al.: Acute stroke: improved nonenhanced CT detection – benefits of softcopy interpretation by using variable window width and center level settings. Radiology 213, 150–155 (1999) 15. Furtado, A.D., et al.: Optimal brain perfusion CT coverage in patients with acute middle cerebral artery stroke. AJNR Am. J. Neuroradiol. 31(4), 691–695 (2010) 16. Brenner, D.J., Hall, E.J.: Computed tomography–an increasing source of radiation exposure. N. Engl. J. Med. 357(22), 2277–2284 (2007) 17. Wintermark, M., et al.: Using 80 kVp versus 120 kVp in perfusion CT measurement of regional cerebral blood flow. AJNR Am. J. Neuroradiol. 21(10), 1881–1884 (2000) 18. Konstas, A.A., Goldmakher, G.V., Lee, T.-Y., Lev, M.H.: Theoretic basis and technical implementations of CT perfusion in acute ischemic stroke, Part 2: technical implementations. AJNR Am. J. Neuroradiol. 30(5), 885–892 (2009) 19. Schad, L.: Sodium imaging revived - clinical and experimental aspects. In: Proceedings of the 28th Annual Meeting of ESMRMB, Leipzig, Germany (2011). https://doi.org/10.1007/ s10334-011-0267-6 20. Mittal, S., Wu, Z., Neelavalli, J., Haacke, E.M.: Susceptibility weighted imaging: technical aspects and clinical applications, Part 2. AJNR Am. J. Neuroradiol. 30(2), 232–252 (2009) 21. Deibler, A.R., Pollock, J.M., Kraft, R.A., Tan, H., Burdette, J.H., Maldjian, J.A.: Arterial spin-labeling in routine clinical practice, Part 1: technique and artifacts. Am. J. Neuroradiol. 29(7), 1228–1234 (2008) 22. Straka, M., Lee, J., Lansberg, M.G., Mlynash, M., Albers, G.W., Bammer, R.: Is reduced CBVA reliable surrogate marker for infarct core and can it be used to identify lesion mismatch? In: Proceedings of the 18th Annual Meeting of ISMRM, Stockholm, Sweden (2010) 23. Kidwell, C.S., et al.: Thrombolytic reversal of acute human cerebral ischemic injury shown by diffusion/perfusion magnetic resonance imaging. Ann. Neurol. 47(4), 462–469 (2000) 24. Olivot, J.M., et al.: Relationships between cerebral perfusion and reversibility of acute diffusion lesions in DEFUSE: insights from RADAR. Stroke 40(5), 1692–1697 (2009) 25. Hemalatha, R.J., Vijaybaskar, V., Dhivya, A.J.A.: Early detection of joint abnormalities from ultrasound images. Int. J. Eng. Technol. (UAE) 7(2.25) (2018) 26. Hemalatha, R, Thamizhvani, T, Dhivya, A.J.A., Joseph, J.E., Babu, B.: Active contour based segmentation techniques for medical image analysis. Med. Biol. Image Anal. 4(17), 2 (2018) 27. Hemalatha, R.J., Vijayabaskar, V.: Histogram based synovitis scoring system in ultrasound images of rheumatoid arthritis. J. Clin. Diagn. Res. 12(8) (2018)
EEG-Based Stress Detection Using K-Means Clustering Method Soumya Samarpita(B) and Rabinarayan Satpathy FOS, Sri Sri University, Cuttack, Odisha, India {soumya.s2020-21ds,rabinarayan.s}@srisriuniversity.edu.in
Abstract. Stress, sadness and panic have all become major issues in our contemporary culture. Stress has become one of the top ten socioeconomic predictors of health inequalities. The electroencephalogram (EEG) signals and machine learning approaches are utilized to predict the mental state of the person. This has become a significant topic of research in recent times in health care system. There are various ways are used to monitor stress. The primary goal of this study is to identify stress in humans. Because of its potential value, stress detection based on EEG signals has emerged as an interesting study topic. This research looks into brain waves to classify a person’s mental state. Despite the fact that there is no precise way of defining the optimum feature for a classifier, the features utilized as classifier input have a significant impact on the classification outcomes. An algorithm for stress level detection from EEG is proposed in this paper. The Euclidean distance scale is commonly used in the paper for EEG signal identification. In this study, EEG data is separated into EEG rhythms using a band pass filter method, EEG signals are normalized and a k-mean clustering method is used to classify brain wave signals to detect the mental stress. Keywords: Brain Waves · Mental Stress · Electro-encephalogram (EEG) · EEG Signals · K-Means clustering
1 Introduction The majority of individuals encounter stress at some point in their lives. Stress can be caused by a variety of factors. It is a common bodily reaction to being challenged or frightened by the surrounding. The body’s response to mental, emotional, or physical discomfort is classified as stress. Stress not only causes unstable behavior, but it may also increase hypertension [17]. Stress is a psychological reaction produced by external stimuli. The two aspects that are followed are mental and physical [15]. Stress can be caused by unfamiliar situations, increasing expectations at work, and emotional reactions to the loss of a loved one or abrupt changes in one’s lifestyle. Due to its potential applications, such as in “Human-Computer Interaction” (HCI), the study of mental stress detection using EEG signals has great promise [8]. The utilization of brainwave signals is a step toward the introduction of individual identification through biometric technology based © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2023 Published by Springer Nature Switzerland AG 2023. All Rights Reserved S. Nandan Mohanty et al. (Eds.): ICISML 2022, LNICST 470, pp. 35–43, 2023. https://doi.org/10.1007/978-3-031-35078-8_4
36
S. Samarpita and R. N. Satpathy
on bodily features. Each individual’s brainwave signal has distinct properties. Periodic stress tracking is required to control the stress experience. It is critical to alleviate human stress; there are various strategies available for doing so, such as yoga and meditation. Different patterns of neuronal interaction in the human brain result in different brain states. The signal waves are distinguished by their amplitudes and frequencies based on these patterns. The neuronal contact occurs between a vast numbers of neurons. EEG is a medium of systematically measuring the electrical activity of central nervous system. These recordings are made on the scalp, and numerous electrodes are inserted in several unique areas on the scalp. Scanning of recorded EEGs assists in defining the condition of the brain with deviations from normal, such as sleep problems, epileptic seizures, mental tension, memory loss etc. on [13]. Different methods, including “PET” (Positron Emission Tomography), “MRI” (Magnetic Resonance Imaging), and “fMRI” (Functional Magnetic Resonance Imaging), can be used to capture brain activity. However, because of its asymptomatic, precision in realtime mode of operation, and optimum suitability for the complicated vitality of human brain processes, EEG signals are the ideal choice to capture the neuronal changes [6]. Medical professionals do not need patient involvement when using EEG-based analysis for diagnostic purposes. In order to distinguish between various types of mental stress based on various variables, this study introduces k-nearest neighbor to be employed as classifiers using the frequency domain feature extraction. We also suggest feature choices to drastically minimize the amount of characteristics, hence reducing the computational complexity. The format of this article is as follows: The comparative analysis is shown in Sect. 2, the proposed methodology for this study is presented in Sect. 3, the experiment’s findings are presented in Sect. 4, and the conclusion is presented in Sect. 5.
2 Comparative Analysis Stress in humans is induced by psychological, intellectual, or bodily resistance to new difficulties. When stress levels surpass a particular threshold, it can have serious consequences for one’s health, mood, creativity, relationships, and life experiences. As a result, early recognition of mental stress is critical for avoiding such negative consequences. EEG microstate analysis is also more effective than resting-state fMRI analysis in terms of its capacity to test expensively big groups of people [16]. The brainwaves signal has various typical and characteristic of the individual, and because brainwaves cannot be duplicated or read by individuals, similarity is not conceivable. Identity identification is required to differentiate an individual’s traits [4]. EEG waves in different mental states include: a) Beta waves (13–35 Hz frequency) b) Alpha waves (8–13 Hz frequency) c) Theta waves (4–8 Hz frequency) and d) Delta waves (0.5–4 Hz frequency). After analyzing various research papers [8, 13], we found that theta waves are mostly produced in the drowsy state, while theta waves are primarily connected with deep sleep. The alpha wave represents a relaxed level of awareness, whereas the beta wave mostly represents a focused state. Gamma waves seldom ever occur.
EEG-Based Stress Detection Using K-Means Clustering Method
37
In this section, the EEG signals from different articles are analyzed with different techniques which are described in Table 1. Table 1. Outline of papers examining how EEG waves affect the human brain. Reference, Years
Aims
Samples
Methods
Conclusions
[14], 2022
“To examine features based on time-domain and frequency-domain for categorising the EEG responses to auditory evoked potentials (AEPs)”
8 healthy participants
As a method for Accuracy:extracting EEG 82.86% characteristics, the Fast Fourier Transform (FFT), power spectral density (PSD), spectral centroids have been used Classification:- SVM, LDA, K-NN
[18], 2022
“To calculate the theta-to-alpha transition frequency from electroencephalographic data in the resting state”
Dataset:OpenNeuro 25 Parkinson’s patients and 25 matched controls
Klimesch’s method K-mean Clustering
Accuracy:- 88% for 2D adjusted k-mean
[7], 2022
“Utilizing audio, video, EEG, and EMG to identify emotions”
Emotion Dataset:CK, CK+ (11 Human Subjects) EEG Dataset:DEAP
Feature Extraction:PCA Classification:- SVM, K-NN,CNN,MLP
Emotion accuracy ranging from 43.4% to 86.1%. Using only eight channels of EEG data and the conventional KNN model, they attained 39.70% accuracy
[3], 2022
“To provide a quick-processing strategy for efficiently extracting and choosing spatial features for emotion recognition”
Dataset:- SEED (15 Chinese participants), DEAP (32 participants)
Features are extracted from DNN Classification:- SVM, K-NN
Accuracy for SEED:- 96.3% (SVM), 86.4% (K-NN) Accuracy for DEAP:- 81.1% (SVM), 79.3% (K-NN)
(continued)
38
S. Samarpita and R. N. Satpathy Table 1. (continued)
Reference, Years
Aims
Samples
Methods
Conclusions
[16], 2022
“To evaluate EEG microstates in MwoA patients during the interictal period”
50 Female participants. “NeuroScan” was used to record the EEG data
Feature Extraction:ICA
Microstate classes B, C, and D were the main areas where microstate syntactic analysis revealed significant differences in transition probabilities between the two groups
[10], 2022
“To suggest a cutting-edge signal processing method for EEG signal emotion identification utilising continuous wavelet transform”
Dataset:- SEED (15 participants)
The BoDF reduces the features Classifiers:- SVM, K-NN
Accuracy for SVM:- 96.7% Accuracy for K-NN:- 95.3%
[1], 2022
“To choose the optimum classification method for the death and surviving cases using COVID19MPD in Mexico”
Dataset:COVID19MPD Samples:200,000patients (30 to 50ages)
Classifiers:- KNN, Naïve Bayes, Random Forest
Accuracy:94.88%
[15], 2021
“To detect stress management using BCI technique”
Raw EEG set
Feature Extraction:Accuracy:Hjorth, KDE, RER, 89.03% for ELC + ELC KNN Classification:- KNN, SVM, NN
[13], 2021
“To examine a discrete wavelet-based feature extraction model for the categorization of EEG signals during a mental arithmetic exercise”
32 participants
Feature Selection:Neighborhood Component Analysis (NCA) Classification:-KNN
Accuracy:- 91%
[2], 2021
“To investigate the electroencephalogram (EEG) data generated when typing”
5 engineering students
Classification:- kNN, Random Forest
Accuracy:98.91% (kNN) 99.89% (RF)
[19], 2021
“To recognize emotions from brain signals for efficient human-robot interaction”
Dataset:- SEED 15 Subjects
Classification:-kNN Optimization:Genetic Algorithm
Accuracy:80.59%
(continued)
EEG-Based Stress Detection Using K-Means Clustering Method
39
Table 1. (continued) Reference, Years
Aims
Samples
Methods
Conclusions
[9], 2018
“To Correlate RAW EEG Signals and Attention Levels of BCI using KNN Technique”
Dataset of 300 mind waves samples
Classification:- kNN
Researchers discovered an excellent relationship between “attention levels” and “RAW EEG” data
[12], 2017
“To evaluate the effectiveness of the K-Nearest Neighbors (K-NN) algorithm for classifying motor imagery using an EEG signal”
4 healthy human subjects
Classification:- SVM, kNN Distance metrics:Manhattan, Euclidean, Minkowski, Chebychev and Hamming
The Minkowski distance has the best classification accuracy (70.08%)
3 Proposed Methodology The technique, suggested model, and how it operates are all included in this section. Additionally, it provides a thorough summary of the feature extraction, classification, and EEG pre-processing. The brain interface method is far superior to other methods because it can analyze stress with a high degree of accuracy using a variety of feature extraction and classifier algorithm combinations. Based on different literature review, our proposed model for the mental stress detection approach utilizing the brain interface technology is illustrated in Fig. 1.
Pre-processing(Removal of Noise)
Feature Extraction(Band-pass filter)
Classification (k-mean) Fig. 1. Stress detection with brain interface techniques
The power per frequency band is the EEG signal characteristic that is most frequently employed for mental stress detection. This suggested method leverages EEG generated by sensors for analysis and is useful for real-time mode stress detection. In this study, multiple temporal and frequency-domain characteristics are used to analyze the EEG
40
S. Samarpita and R. N. Satpathy
signals, and the recovered features are then categorized using the K-means clustering method. In this regard, a bandpass filter with a range from 0.5 to 60 Hz is applied for EEG classification. The experiment in this paper uses four bands (BETA, ALPHA, THETA, and DELTA). The frequency domain properties are used to compute band power for these four bands. The K-means [5], simplest unsupervised clustering technique, which divides each item into a cluster based on the observation’s mean. K-means typically moves from one cluster to another in finding of the k-partition calculating within cluster sum of square. K-means clustering may divide an input set of n parameters into an equal number of clusters, with each observation belonging to the cluster that is closest to the centroids, which act as prototypes for the clusters. By reducing the Euclidean distance between the “data” and the matching “cluster centroid”, data are clustered. K-mean clustering divides the data into ‘k’ clusters in this manner [11]. K-Means is a clustering technique that categorizes data into k groups. Because we may choose the number of clusters, it is easily applicable in classification, where we split data into clusters that are equal to or more than the number of classes. K-Means Algorithm:Step 1: First, determine the number of clusters ‘K.’ Step 2: As centroids for each cluster, randomly select K data points. The value of ‘K’ will be 2 if there are two clusters. Step 3: Iterate numerous times until the data points allocated to clusters do not change. Step 4: Sum the squared distances between the data points and the centroids. Step 5: To minimise the distance, assign each data point to the nearest cluster (centroid). Step 6: Take an average of the centroids of the clusters that are related. This is a one-time operation that computes the centroid and assigns points to the cluster depending on their distance from the centroid. The procedure is terminated once all centroids have been established. In this paper, Euclidean distance is calculated to find the mean. The k-means clustering technique measures the similarity of items using the Euclidean distance. For typical k-means clustering, both iterative and adaptive algorithms are available. K-means clustering techniques must make the assumption that the number of groups (clusters) is known in advance. The Euclidean distance is as follows: N (mi − ni )2 (1) dis(m, n) = i=1
where, dis = distance from point m to point n, N = N- sample space and m, n = two points, mi , ni = Euclidean vectors are drawn from the origin of space (initial point)
4 Results The data extracted from the characteristic findings are normalized and grouped into four groups using K-Means Clustering. These four groups are alpha, beta, theta and delta waves respectively. Then these waves are converted to corresponding EEG signals.
EEG-Based Stress Detection Using K-Means Clustering Method
41
These EEG signals are normalized using band-pass filter method to get the features like frequency and phase. After normalization, frequencies vs. phase points are coordinated to form the cluster. Figures 2 shows the frequency of the alpha wave, beta wave, theta wave and delta wave plotted with phase in degree.
Fig. 2. Frequency of alpha, beta, theta and delta Wave
The findings are achieved using Matlab to execute the K-Mean approach. In this study, 32 samples are taken to classify the EEG signals. Here, only four clusters C1, C2, C3 and C4 are taken into consideration and the data matrix is grouped using the k-means clustering method. These four clusters represent the mental state of the person. The C1 cluster indicates that the subject is sleeping deeply (delta wave). The C2 cluster indicates that a person is extremely calm (theta wave). The C3 cluster denotes when the individual is in a highly calm condition (alpha wave), whereas the C4 cluster reflects the beta wave, which means the person is in an active or relaxed state. Table 2. Output result of Stress Analysis.
Here Table 2 depicts a person’s stress analysis.. Four clusters are classified as being in various states of stress because we are considering 32 samples and each cluster has 8. From these points, C1 shows the delta wave (0 Hz–4 Hz), C2 represents the theta wave (4 Hz–8 Hz), C3 represents the alpha wave (8 Hz–16 Hz), and C4 shows the beta wave (16 Hz–35 Hz).
42
S. Samarpita and R. N. Satpathy
The summaries of different clusters that represent stressed or relaxed condition of the person are depicted in the Fig. 3.
Fig. 3. Clustering with K-means
5 Conclusion The EEG signal is an electrical signal produced by coordinated brain activity. EEG may be used to detect anomalies in brain waves and classify distinct mental states. Classifiers are used to classify EEG signals in order to detect anomalies in brain waves. This suggested classification approach is distinct and makes it extremely simple to identify EEG data. In this paper, a real-time EEG-based stress detection algorithm is used. The K-Mean clustering method is used to produce four stages of stress and EEG data is used to check the suggested stress detection system. The K-mean clustering method is also used to investigate frequency characteristics. In the future, it is intended to collect EEG data from the participants and a database can be created which will be used for EEGbased stress identification. And, different feature extraction and classification techniques could be analyzed for further study.
References 1. Almustafa, K.M.: Covid19-Mexican-patients’ dataset (Covid19MPD) classification and prediction using feature importance. Concur. Comput. Pract. Exp. 34(4), e6675 (2022) 2. Amalina, I., Saidatul, A., Fook, C.Y.: Frequency bands based on EEG typing for biometric authentication. In: AIP Conference Proceedings, vol. 2339, no. 1, p. 020170. AIP Publishing LLC, May 2021 3. Asghar, M.A., Khan, M.J., Rizwan, M., Shorfuzzaman, M., Mehmood, R.M.: AI inspired EEG-based spatial feature selection method using multivariate empirical mode decomposition for emotion classification. Multimedia Syst. 28(4), 1275–1288 (2022) 4. Azhari, A., Hernandez, L.: Brainwaves feature classification by applying K-means clustering using single-sensor EEG. Int. J. Adv. Intell. Inform. 2(3), 167–173 (2016)
EEG-Based Stress Detection Using K-Means Clustering Method
43
5. Bablani, A., Edla, D.R., Kuppili, V., Ramesh, D.: A multi stage EEG data classification using k-means and feed forward neural network. Clin. Epidemiol. Glob. Health 8(3), 718–724 (2020) 6. Bavkar, S., Iyer, B., Deosarkar, S.: Detection of alcoholism: an EEG hybrid features and ensemble subspace K-NN based approach. In: Fahrnberger, G., Gopinathan, S., Parida, L. (eds.) ICDCIT 2019. LNCS, vol. 11319, pp. 161–168. Springer, Cham (2019). https://doi. org/10.1007/978-3-030-05366-6_13 7. Chen, J., Ro, T., Zhu, Z.: Emotion recognition with audio, video, EEG, and EMG: a dataset and baseline approaches. IEEE Access 10, 13229–13242 (2022) 8. Diah, K.T., Faqih, A., Kusumoputro, B.: Exploring the feature selection of the EEG signal time and frequency domain features for k-NN and weighted k-NN. In: 2019 IEEE R10 Humanitarian Technology Conference (R10-HTC) (47129), pp. 196–199. IEEE, November 2019 9. Garg, P., Singh, R.P., Mehra, M.: To correlate RAW EEG signals and attention levels of BCI using KNN technique (2018) 10. Haq, Q.M.U., Yao, L., Rahmaniar, W., Islam, F.: A hybrid hand-crafted and deep neural spatio-temporal EEG features clustering framework for precise emotional status recognition. Sensors 22(14), 5158 (2022) 11. Hegde, N.N., Nagananda, M.S., Harsha, M.: EEG signal classification using k-means and fuzzy c means clustering methods. Int. J. Sci. Technol. Eng 2(1), 1–5 (2015) 12. Isa, N.E.Z.M., Amir, A., Ilyas, M.Z., Razalli, M.S.: The performance analysis of Knearest neighbors (K-NN) algorithm for motor imagery classification based on EEG signal. In: MATEC Web of Conferences, vol. 140, p. 01024. EDP Sciences (2017) 13. Islam, A., Sarkar, A.K., Ghosh, T.: EEG signal classification for mental stress during arithmetic task using wavelet transformation and statistical features. In: 2021 International Conference on Automation, Control and Mechatronics for Industry 4.0 (ACMI), pp. 1–6. IEEE, July 2021 14. Islam, M.N., Sulaiman, N., Rashid, M., Mustafa, M., Jadin, M.: Investigation of time-domain and frequency-domain based features to classify the EEG auditory evoked potentials (AEPs) responses. In: Ab. Nasir, A.F., Ibrahim, A.N., Ishak, I., Mat Yahya, N., Zakaria, M.A., P. P. Abdul Majeed, A. (eds.) Recent Trends in Mechatronics Towards Industry 4.0. LNEE, vol. 730, pp. 497–508. Springer, Singapore (2022). https://doi.org/10.1007/978-981-33-45973_45 15. Lahane, P.: Brain computer interfaces techniques for stress management. Inf. Technol. Ind. 9(3), 767–775 (2021) 16. Li, Y., et al.: Abnormalities in resting-state EEG microstates are a vulnerability marker of migraine. J. Headache Pain 23(1), 1–12 (2022). https://doi.org/10.1186/s10194-022-01414-y 17. Shon, D., Im, K., Park, J.H., Lim, D.S., Jang, B., Kim, J.M.: Emotional stress state detection using genetic algorithm-based feature selection on EEG signals. Int. J. Environ. Res. Public Health 15(11), 2461 (2018) 18. Vallarino, E., Sommariva, S., Famà, F., Piana, M., Nobili, F., Arnaldi, D.: Transfreq: A Python Package for Computing the Theta-to-Alpha Transition Frequency From Resting State Electroencephalographic Data. Wiley, Hoboken (2022) 19. Vrochidou, E., Lytridis, C., Bazinas, C., Papakostas, G.A., Wagatsuma, H., Kaburlasos, V.G.: Brain signals classification based on fuzzy lattice reasoning. Mathematics 9(9), 1063 (2021)
Detection of Psychological Stability Status Using Machine Learning Algorithms Manoranjan Dash1(B) , M. Narayana2 , Nampelly Pavan Kalyan2 , Md Azam Pasha2 , and D. Chandraprakash3 1 Artificial Intelligence Department, Anurag University, Hyderabad, India
[email protected]
2 Electronics and Communication Engineering Department, Anurag University, Hyderabad,
India [email protected] 3 Electronics and Communication Engineering, KG Reddy College of Engineering and Technology, Hyderabad, India [email protected]
Abstract. Obviously, individuals all over the world make a solid effort to stay aware of the hustling scene. Nonetheless, thus, every man and lady is managing interesting wellness issues, one of the most notable of which is misery or stress, which can prompt passing or other horrifying demonstrations. These inconsistencies are alluded to as bipolar problem, which can be treated by following a couple of expert suggested medicines. Victims who have been determined to have psychological wellness issues have their circumstances analyzed to assist them with approaching their regular routines. Positive conditions, such as Schizophrenia and Bipolar Disorder, have a higher likelihood of continuing crises. Mental health professionals are responsible for reducing the risk of patients experiencing crises. Machine learning is being used by neuroscientists and therapists all around the world to widen treatment regimens for patients and to identify some of the key signs for mental health issues before they manifest. One of the benefits is that device learning helps practitioners to predict who might be at risk of a specific condition. For this study, statistics were gathered from working humans, and the dataset was ran through a few machine mastering algorithms, which included all forms of queries for depressed identification. When compared to DNN and Logistic Regression, the Random Forest algorithm delivers the best accuracy of 81.02% after applying a few algorithms to the data set. Keywords: Machine Learning · Logistic Regression · Random Forest · Deep Neural Network
1 Introduction Emotional well-being can affect day to day existence, relatives, and actual wellbeing. Regardless, this association additionally works the other way (Azar et al. 2015). Factors in individuals’ encounters, social affiliations, and actual attributes could be in every way © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2023 Published by Springer Nature Switzerland AG 2023. All Rights Reserved S. Nandan Mohanty et al. (Eds.): ICISML 2022, LNICST 470, pp. 44–51, 2023. https://doi.org/10.1007/978-3-031-35078-8_5
Detection of Psychological Stability Status
45
transferred to cause psychological well-being issues. Dealing with emotional well-being issues can tremendously affect an individual’s point of view. This will help you in arriving at life agreement. Stress, gloom, and stress, for instance, can all impact mental wellness and disrupt an individual’s everyday movement (Ranjana et al. 2019). Regardless of the way that the terms emotional well-being and psychological sickness are much of the time utilized conversely, a few circumstances that specialists view as mental challenges have actual roots. Monetary elements, for example, whether occupation is open in the area, profession, and an individual’s degree of social thought, training are modifiable factors for emotional wellness issues, while non-modifiable factors incorporate orientation and age (Saha et al. 2016). Psychological sicknesses are much of the time connected to various actual infirmities, like coronary supply route infection and diabetes. The treatment of scholarly infirmity has been founded on the conviction that issues of feeling, addressing, and direct don’t be guaranteed to require realness and on second thought address interesting blemishes or unfortunate way of life decisions (Simms et al. 2017). Most emergency work environments are ill-equipped to manage patients’ concerns during psychological wellness rise. Most insurance programs view psychological sickness and compulsion as exemptions for the overall rule, as opposed to a piece of it (Ahmad et al. 2017). No matter what a famous social pattern toward compassion, our overall local area will see the intellectually debilitated and those with proclivity as ethically harmed as opposed to exhausted while everything is said and done (Sumathi and Poorna 2016). Regarding the matter of forecast, researchers have utilized and explored various strategies. Wrongly treating psychological sickness can bring about irreversible harm to the patient’s emotional wellness and even demise (Sauter et al. 1999). Countless enduring all over the planet are not as expected really focused on. In this exploration, a solitary report lays out a semi-robotized structure that guides in the underlying responsibility of the psychological well-being patient (Leung et al. 2008). Psychological maladjustment essentially affects every relative, as well as the man or lady and society. Relational gatherings empower people with mental sickness to speak with other people who are additionally impacted by dysfunctional behavior through webbased correspondence, giving data about psychological maladjustment issues (Aricò et al. 2017). Diseases of the psyche habitually happen in gatherings; for instance, an individual experiencing stress may likewise encounter melancholy. The converging of scholarly circumstances focuses on our work of sorting out web networks with a propensity for torment. In this paper, dysfunctional behavior issues have turned into a significant issue in the public eye, and it likewise affects a person’s regular repetitive work. Numerous medical conditions emerge because of stress and gloom. In this situation, a goal measure for recognizing the phases of pressure while considering the brain ought to extraordinarily expand the connected adverse consequences. As a result, an AI shape protected by an EEG sign is constructed in this study. The quit impacts help to make sense of why the trend setting innovation has a 95% precision rate. The predetermined EEG structure gives a layered pressure objective that might be measured. It can likewise be utilized to make a programmed pressure recognition apparatus.
46
M. Dash et al.
Precautionary determination of cerebrum tainting can help with more viable treatment and further develop the singular’s general endurance possibilities. It is basic to manage such issues quickly to stay away from death toll. Man-made intelligence and AI approaches can be utilized to analyze and treat an assortment of medical issues. They took and utilized seven gadget learning calculations to find precision for five wellness related issues in this examination. For the interaction, a bunch of information with 59 events is utilized. The calculations were all sudden spike in demand for the dataset and yielded an elevated degree of exactness with the littlest conceivable variant.
2 Problem Formulation After India’s populace blast, the proportion of specialists to patients is 1: 1810, and an expert’s time enjoyed with a patient is under minutes. Despair is the most wellknown reason for global insufficiency (Graziani et al. 2016). In every practical sense, the vast majority of individuals with mind issues are misdiagnosed in overall regions, with roughly 1,000,000 individuals ending their own lives every year. Likewise, as indicated by WHO research, 1 out of each and every 14 individuals had an uncomfortable inclination (Kawakami et al. 1995). As indicated by the World Wellbeing Association, stress concerns are the most usually seen mental issues around the world, with explicit fear, significant extreme issue, and social fear being the main uneasiness issues. Sridharan et al. (2015) used Convolution Neural Networks (CNN) to give discovery diagnostics on web-based virtual entertainment, with the accentuation being on getting insights revealed by different shoppers while likewise guaranteeing that a bunch of rules protects the security by isolating merchants who manage realities. On the clinical dataset produced, the Stollar et al. (2010) N B Allen approach utilizes upgraded unearthly pass off boundaries for identification of burdensome incidental effects from discussion signals. The characterization of these features is finished with the assistance of a straightforward SVM classifier. Orientation reliance has advanced despairing sort both great for women, guys, and changed in the midst of features in past exploration. In this test, grown-up guys had a preferable possibility distinguishing despair over females. Ang Li et al. (2016) directed a language study to distinguish sadness. Patricia A. et al. Directed a substance investigation of horror related Tweets. A broadly specialist examination among U.S. More youthful grown-ups become wrapped up by Brian et al. utilizing explicit online entertainment stages (Morris 1995). Emotional component examination of Online Melancholy People group transformed into finished through Thin Nguyen et al. (2018) the significant motivation behind those designs is to structure the calculation for identification of the slump shame gainfully. To start with, the assembled information is examined utilizing punctuation and semantics assessment, which brings about the impression of a distress shame among postings made by individuals of different ages. In this system, the language structure is analyzed to recognize explicit watchwords, and the pertinence of these watchwords is resolved utilizing semantic assessment, which uncovers the general feeling of the passage by dissecting the substance material’s inclination, otherwise called Feeling Discovery Frameworks. By then, the presents are named in view of the results of the descending perspective. Late works have zeroed in on the decay of substance because of virtual entertainment, but the functioning individuals’ pressure has been ignored.
Detection of Psychological Stability Status
47
3 Methodology The proposed gadget considers a portion of the tech individuals’ strain identification. The dataset under assessment is a review of various working staff, which incorporated all potential strain discovery questions. For stress distinguishing proof, the planned strategy utilizes the ML calculation; for dominating and recognition, the dataset is utilized with Arbitrary Lush Region and Calculated Relapse. The proposed technique gives an exact arrangement of standards for foreseeing dysfunctional behaviors. The code was written in Python, and fundamental libraries were utilized (Packhauser et al. 2022). The information was gotten by means of kaggle. The information is then isolated into two gatherings: preparing and looking at. The utilization of AI calculations that might be appropriate for this issue is made.
Fig. 1. Flow process of proposed model
Figure 1 shows the point by point stream interaction of the proposed model. In the initial step, different information with regardless age and orientation has been gathered. After the information assortment, some minor pre-handling is done to change over the information into reasonable handling design. Various highlights are removed and the ideal contributed highlights are held prior to applying the AI strategies. Three AI procedures to be specific Irregular Woods, DNN and Strategic Relapse have been carried out and checked with preparing and testing stage. At long last the quantitative execution examination in wording precision is figured.
4 Results and Discussion 4.1 Dataset • It is a review dataset that explores human conduct as far as scholarly contaminate particle and the recurrence of scholarly issues in programming workplaces. The accompanying credits make up the dataset: • Age, Orientation, and Nation are completely remembered for the timestamp. • State: where do you plan to remain? • Independent work: Yes or No? • Family ancestry: Do you have a circle of family members who have a past filled with psychological sickness? • Treatment: Have you attempted any emotional well-being medicines? • Work meddling: Do you observe that your scholarly wellbeing is interfering with your work? • Number of representatives: What number of laborers do you have in your office? • Remote work: Do you telecommute or somewhere else?
48
M. Dash et al.
• Number eight: Is it important to have a specialized association? • Benefits #9: Does your working environment give psychological well-being benefits? • Choices to mind: Would you say you are mindful of your manager’s possibilities for scholarly human administrations? Health program: Has your boss referenced mental prosperity as a feature of a decent workout schedule anytime? • Seek assistance: Does your boss allow space to observe mental health issues and a mechanism to seek help? • Thirteen. Is your anonymity guaranteed if you use mental health or substance abuse treatment resources? • Leave: Is it simple to take clinical leave for an intellectual health problem? • Mental health effect: Do you believe that discussing a mental health issue with your boss would have negative consequences? • Physical fitness result: Do you believe that discussing a physical clinical issue with your supervisor would have negative consequences? • Colleagues: Would you enjoy studying an intellectual medical problem with your coworkers? • Supervisor: Would you be willing to discuss a mental scientific issue with your immediate supervisor(s)? • Will you create a mental clinical hassle with a potential supervisor in an assembly if you have a mental fitness interview? • Physical health interview: Will you bring up a physical • Mental health vs. physical fitness: Do you think your boss is as concerned about your mental health as he or she is about your physical well-being? • Obs effect: Have you seen or heard about negative consequences for collaborators with mental health issues in your running environment? • Remarks or Observations: Do you have any more thoughts? 4.2 Quantitative Analysis The framework is further developed on the grounds that it is written in Python and incorporates the important libraries. Irregular forested region rendition beats unmistakable models while carried out using three contraption dominating strategies on the provided dataset for mental infection recognizable proof. When contrasted with elective Calculated Relapse sets of rules, Arbitrary Backwoods and DNN calculations show great exactness. Table 1. Experimental Results of Proposed System Algorithm
Accuracy
Random Forest
81.22
DNN
80.42
Logistic Regression
79.37
The exact recognition of mental dependability as far as precision metric utilizing three different AI calculations to be specific Irregular Woods, DNN and Strategic Relapse is
Detection of Psychological Stability Status
49
organized in Table 1. From Table 1. It tends to be seen that the precision is ideal in the event of Arbitrary Backwoods when contrasted with DNN and Strategic Relapse AI procedures.
Fig. 2. Correlation Matrix
Figure 2 shows the relationship network of the proposed model. Connection is a sign about the alterations among factors. Connection grid shows which variable is having an unnecessary or low relationship in respect to each and every other variable.
Fig. 3. Bar plot indicates probability of mental health condition with age
50
M. Dash et al.
Figure 3 shows bar plot that indicates probability of mental health condition with age. It can be observed from the figure that the age group above 60 years have the high probability of mental disorder.
5 Conclusion There are different methodology that can be utilized to recognize scholarly tainting in individuals of different ages. These frameworks utilize a methodology of recognition that includes dissecting the psychological issue identification involving a progression of inquiries to foresee the phases of decline across different age gatherings. For recognizing scholarly disarray, framework dominating calculations are utilized. The dataset, which contains 1257 examples, is being inspected. For dominating and discovery, we utilized Irregular Woodland, DNN, and Strategic Relapse. The Irregular Backwoods accomplishes the most noteworthy exactness of generally 81.02%, as indicated by the trial information. The exactness of the DNN calculation is around 80%, while the exactness of Strategic Relapse is around 80%. In the future, we should embrace legitimate pre-handling of the informational collection to accomplish higher correctnesses, and we should utilize further developed AI calculations. We can accomplish top precision utilizing an exchange approach, which involves consolidating astounding gadget learning calculations to accomplish higher exactnesses.
References Azar, G., Gloster, C., ElBathy, N., Yu, S., Neela, R., Alothman, I.: Intelligent data mining and machine learning for mental health diagnosis using genetic algorithm, pp. 201–206 (2015). https://doi.org/10.1109/EIT.2015.7293425 Mental disorder detection: bipolar disorder scrutinization using machine learning. Int. J. Adv. Comput. Sci. Appl. (2019). https://doi.org/10.14569/IJACSA.2019.070170 Saha, B., Nguyen, T., Phung, D., Venkatesh, S.: A framework for classifying online mental healthrelated communities with an interest in depression. IEEE J. Biomed. Health Inform. 20(4), 1008–1015 (2016) Simms, T., Ramstedt, C., Rich, M., Richards, M., Martinez, T., Giraud-Carrier, C.: Detecting cognitive distortions through machine learning text analytics. In: 2017 IEEE International Conference on Healthcare Informatics (ICHI), Park City, UT, pp. 508–512 (2017) Subhani, A., Mumtaz, W., Saad, M., Naufal, M., Kamel, N., Malik, A.: Machine learning framework for the detection of mental stress at multiple levels. IEEE Access, 1 (2017). https://doi. org/10.1109/ACCESS.2017.2723622 Sumathi, M.R., Poorna, B.: Prediction of mental health problems among children using machine learning techniques. Int. J. Adv. Comput. Sci. Appl. (2016). https://doi.org/10.14569/IJACSA. 2016.070176 Sauter, S., et al.: Stress at Work, DHHS (NIOSH) Publication No. 99-101. NIOSH Cin Cinnati (1999) Leung, M.Y., Chan, Y.S., Olomolaiye, P.: Impact of stress on the performance of construction project managers. J. Constr. Eng. Manag. 134(8), 644–652 (2008) Aricò, P., et al.: Human factors and neurophysiological metrics in air traffic control: a critical review. IEEE Rev. Biomed. Eng. 10, 250–263 (2017)
Detection of Psychological Stability Status
51
Graziani, I., et al.: Development of the human performance envelope concept for cockpit HMI design. In: HCI-Aero 2016 International Conference on Human- Computer Interaction in Aerospace (2016) Kawakami, N., Kobayashi, F., Araki, S., Haratani, T., Furui, H.: Assessment of job stress dimensions based on the job demands- control model of employees of telecommunication and electric power companies in Japan: reliability and validity of the Japanese version of the job content questionnaire. Int. J. Behav. Med. 2(4), 358–375 (1995) Mucci, N., et al.: Work-related stress assessment in a population of Italian workers. The stress questionnaire. Sci. Total Environ. 502, 673–679 (2015) Bashir, U., Ismail Ramay, M.: Impact of stress on employees job performance: a study on banking sector of Pakistan (2010) Leung, M.Y., Liang, Q., Chan, I.Y.: Development of a stressors– stress– performance–outcome model for expatriate construction professionals. J. Constr. Eng. Manag., 04016121 (2016) Morris, J.D.: Observations: SAM: the self-assessment manikin; an efficient cross-cultural measurement of emotional response. J. Advert. Res. 35(6), 63–68 (1995) Tsutsumi, A., Shimazu, A., Eguchi, H., Inoue, A., Kawakami, N.: A Japanese stress check program screening tool predicts employee longterm sickness absence: a prospective study. J. Occup. Health 60(1), 55–63 (2018) Packhauser, K., et al.: Deep learning-based patient re-identification is able to exploit the biometric nature of medical chest Xray data. Sci. Rep. 12(1), 1–13 (2022)
GLCM Based Feature Extraction and Medical X-ray Image Classification Using Machine Learning Techniques Jyotiranjan Rout(B) , Swagat Kumar Das, Priyabrata Mohalik, Subhashree Mohanty, Chandan Kumar Mohanty, and Susil Kumar Behera Department of Computer Science and Engineering, Balasore College of Engineering and Technology, Balasore, India [email protected]
Abstract. The automated system is now created with excellent accuracy to detect abnormalities in X-ray images. To enhance the appearance of medical photographs, image pre-processing methods are applied, so that high accuracy can be achieved with constrained means. Images are often classified based on their textural properties, which are measured using the Gray Level Co-occurrence Matrix (GLCM). The grey level correlation matrix provides statistical information of the second order on the grey levels of neighboring pixels in a picture (GLCM). In this proposed paper, medical X-ray images are classified and their features are extracted using an ensemble learning model. By extracting image features using the GLCM feature extraction method, this proposed model is able to distinguish between healthy and sick images (Gray level co-occurrence matrix).to improve the efficiency of the Ensemble learning classification method, it is compared against various algorithms using performance indicators, including Logistic regression, Gaussian Naive Bayes, as well as Random Forest. When this approach is compared to existing methods, the proposed ensemble model has an accuracy rate of 97% in classifying normal and diseased images. Keywords: GLCM · Machine Learning · random forest · SMOTE
1 Introduction The field of clinical imaging will likely find applications for machine learning as well as artificial intelligence in the not-too-distant future. A number of diagnostic tools, including X-ray machines, magnetic resonance imaging (MRI) scanners, as well as computed tomography (CT) scanners, assist medical professionals in recognizing problems in clinical images more rapidly [1]. Some of the more complex characteristics are utilized to identify things that regular people are incapable of recognizing without the assistance of machine learning as well as artificial intelligence. This new era, marked by the rise of intelligent systems that provide assistance to machines, offers tremendous assistance to radiologists in clinical processes. © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2023 Published by Springer Nature Switzerland AG 2023. All Rights Reserved S. Nandan Mohanty et al. (Eds.): ICISML 2022, LNICST 470, pp. 52–63, 2023. https://doi.org/10.1007/978-3-031-35078-8_6
GLCM Based Feature Extraction and Medical X-ray Image Classification
53
Image processing methods play an essential role in medical image interpretation and automated diagnosis. In recent years, the development of whole-slide imaging methods and the rise in cancer cases have attracted the interest of several researchers in image processing. The whole slide photographs have a very high quality, and professionals need a lot of time to weigh in on them [2]. Automated image processing approaches helped by computers are offered to facilitate this exhaustive procedure. These approaches assist the expert in determining the image analysis, and in certain instances take the position of decision maker. The automated categorization of pictures and the decision-making process rely on visual characteristics. Numerous visual characteristics, such as texture variances, shape discrepancies, light fluctuations, and color shifts, offer categorization systems with important information. The most crucial aspect here is to find the right characteristics and classification method for these features. Various feature extraction techniques may provide different categorization outcomes for the same picture. Consequently, feature selection is one of the most crucial steps in classification. Clinical imaging’s future applications will include machine learning and artificial intelligence. Some equipment, X-rays, MRI, and CT scans are just a few examples of clinical imaging modalities that help clinicians spot anomalies quickly and accurately. Some of the complex qualities are put to use in order to unearth abnormalities machine learning and AI are required to see this. The radiologist analyses X-ray pictures for bone fracture detection using their training and experience [3]. Automatic anomaly identification of X-ray images assists radiologists in diagnosing a variety of clinical conditions, including arthritis, dental caries, bone cancer, osteoporosis, fracture, as well as infection. Examining medical pictures is a complex process that calls for a physician’s full arsenal of skills, expertise, as well as diagnostic equipment. Clinical X-ray picture anomaly identification via an automatic technique is a challenging problem in machine learning. Patients have wide anatomical variants. As a result, this is a major obstacle when trying to project radiographs that have structures superimposed on top of them [4]. As it is difficult for the medical professional to spot the anomaly due to the low quality of the clinical picture taken by the image capture equipment, a suitable computer-assisted automated detection approach is proposed. Consequently, the radiologist may benefit from automated anomaly identification of X-ray images in the diagnosis of a wide range of clinical issues, including arthritis, dental caries, malignant neoplasm, osteoporosis, fracture, as well as infection. 1.1 Computer-Aided Detection (CAD) Systems Medical image analysis is a complex procedure that calls for a physician’s substantial training, experience, and imaging equipment. It is a challenging field of machine learning to develop an automated method for detecting abnormalities in clinical X-ray pictures. The goal of the system is to improve diagnostic accuracy. There is a wide range of anatomical possibilities for each individual patient [4]. As a result, this is one of the most important concerns of radiograph projection, which deals with structures that are overlaid on one another. When fracture detection by X-ray analysis, the radiologist relies on their expertise as well as their experience in the field. Medical professionals may miss an anomaly due to poor image quality from the image capturing device. An efficient automated pc-aided detecting approach has been developed to address this problem.
54
J. Rout et al.
Therefore, many clinical issues, including arthritis, tooth decay, malignant neoplasm, osteoporosis, fracture, as well as infection, may be identified by the radiologist with the use of automated abnormality detection of an X-ray image. These are only few of the conditions that may be diagnosed. 1.2 Musculoskeletal Radiographs (MURA) X-rays of the musculoskeletal system are compiled into a huge dataset known as a musculoskeletal radiograph (MURA). To determine if X-ray scans should be considered normal or abnormal, the calculations have to be believed. More than 30 million individuals go to the emergency department every year due to musculoskeletal difficulties; these issues impact upwards of two billion people worldwide and are the major cause of serious, long-term pain and impairment. It is possible that groundbreaking discoveries in medical imaging will be sparked by the MURA dataset, which in turn may analyze the contribution of professionals to the process of social insurance reform. The Medical Image Archive and Registry (MURA) is a major resource for open radiography research that comprises both normal and abnormal X-ray pictures. There are a total of 36770 pictures, 20828 of which are normal X-rays and 15942 of which are abnormal, or 56.64% normal data as well as 43.36% aberrant data, respectively. 1.3 Contribution of the Study The major focus of this study is on the use of GLCM for feature extraction and ensemble learning for medical X-ray picture categorization. The study explores the possibility of using an ensemble model to detect the normal as well as diseased image. 1.4 Objectives of the Study The following are the objectives that need to be accomplished in order to do this study. To model an ensemble learning that can be used to detect the normal image and diseased image. To analyze the performance of ensemble learning to classify the images. 1.5 Organization of the Paper The remainder of paper is structured as follows, In Sect. 2 the recent work in the medical X-ray image classification and GLCM based features are presented. In Sect. 3 the proposed methodology is clearly described. In Sect. 4 the obtained results are presented with a brief discussion. And finally, the paper is concluded in Sect. 5.
GLCM Based Feature Extraction and Medical X-ray Image Classification
55
2 Literature Review J. Ding et al. [5] in 2017 we put into place a variety of support vector machines, K-means clustering, as well as a deep learning approach are only few of the machine learningbased techniques. Seventeen GLCM characteristics were used to define these methods, and these were extracted for use in testing and vetting terabytes of image data. In the year 2018, C. Xia et al. [6] performed using a support vector machine to identify bone tumors from X-ray images based on their texture, as determined by a 5-fold crossvalidation procedure. This was done by identifying bone tumors using GLCM-based texture features. In the year 2019, Z. Xing et al. [7] suggested an enhanced salap swarm technique for optimizing GLCM using a fitness function as well as diagonal class entropy. This algorithm was tested on natural satellite images. Many of the authors discussed related to medical X-ray image classification and feature extraction by using different methods and algorithms. Some of them are discussed below with accuracy values (Table 1). Table 1. Summarization of literature review S. NO
AUTHOR
YEAR
PROPOSED METHODOLOGY
RESULT(ACCURACY)
[8]
Pawan Kumar Mall et al.
2019
For the purpose of anomaly identification, LBF SVM (Radial Basis Functions support vector machine), linear SVM (standard), logistic regression (regression with intermediate steps), and decision tree are the four classifiers to choose from
The classification accuracy of the LBF SVM classifier was 62.00%. However, LBF SVM outperforms the other three models significantly, This, considering the visual challenges of identifying factures in the MURA database, is cause for optimism
[9]
Qingchen Zhang et al.
2020
For medical image analytics, a graphics processing unit (GPU) is essential in the field of “smart medicine.”
The provided approach improves medical picture classification performance without reducing efficiency compared to conventional architecture and VGG-16
[10]
Basra Jehangir et al.
2022
Chest X-ray images are analyzed using a GLCM-based LGBM classification for the COVID-19 study
The suggested technique was verified by comparing it to the COVID-19 X-ray dataset, as well as an accuracy of 92.40% was obtained
[11]
C.M.A.K. Zeelan Basha
2018
An artificial neural network-based strategy for extracting and classifying statistical features is described
The findings demonstrate that the provided method is an efficient tool for categorising X-ray pictures, with an overall accuracy of 92.3%
3 Methodology The major goal of this study was to create a technique for extracting characteristics from the medical X-ray image and then classifying those features using ensemble learning. Using a feature extraction technique called GLCM, we were able to distinguish between
56
J. Rout et al.
healthy and sick photos in this research (Gray level co-occurrence matrix). GLCM characteristics that were analyzed in this research are contrast, correlation, dissimilarity, energy, homogeneity. In this study, there are two classes to examine, those are described below (Figs. 1 and 2): • Normal • Disease
Fig. 1. From the above figure, it is observable that the count of each class is different data.
Fig. 2. Counts after SMOTE
For this, we used SMOTE analysis to adjust each and every class with same data. Finally, Performance measurements are used to compare the Ensemble learning classification algorithm against other algorithms like Logistic regression, Gaussian Naive Bays, as well as Random Forest algorithms in order to demonstrate its effectiveness.
GLCM Based Feature Extraction and Medical X-ray Image Classification
57
3.1 SMOTE Analysis SMOTE stands for Synthetic Minority Oversampling Technique. SMOTE is an oversampling strategy that involves creating fabricated samples for the underrepresented group. This strategy is useful for avoiding the over fitting problem that might arise from using a large sample size at random. It focuses on the feature space to generate new instances by interpolating between nearby positive examples. Feature Extraction The process of extracting features from a picture is an essential component of the model that we have suggested. The co-occurrence of neighboring grey levels as well as the counts of those levels in a picture are the basis for the GLCM texture feature, which operates on this phenomenon. The ROI (region of interest) dimension of the a matrix used to compute the GLCM texture feature is set to equal the number of grey levels (N) there in X-ray pictures. 3.1.1 GLCM The textural characteristic known as GLCM (Gray level co-occurrence matrix) relies on the frequency with which neighboring grey levels appear in a given image. Depending on the number of grayscale values, a square matrix is used to calculate a GLCM texture feature (N) in the area of interest of the X-ray pictures. GLCM is a prominent approach for extracting texture-based features. This technique uses two metrics to synthesize image features. The GLCM features are calculated during the first step, and the texture properties based on the GLCM are derived during the second process. For a couple of the novel characteristics, the metric formula is illustrated below. The feature is extracted using a grey level co-occurrence matrix. This investigation made use of the following GLCM characteristics; contrast, correlation, dissimilarity, energy, homogeneity. 3.1.2 GLCM Features Contrast: A metric for determining how often certain parts of an image appear, or the intensity of a pixel and its surrounding pixels, is called contrast. It also establishes how many local differences are visible in the picture. The contrast is calculated by comparing the hue and luminosity levels of the target object to those of other objects in the same visual field. The following equation characterizes the contrast function of any given image: |i − j|2 p(i, j) (1) cont. = i
j
Correlation: A pixel’s association with its neighboring pixels throughout the whole picture is measured by an image characteristic called correlation. The correlation coefficient ranges from −1 to 1 for properly positively correlated images, and it is infinite for constant images. Equation defines the word “Correl” as the correlation characteristic of a picture. (i − μi)(j − μj)p(i, j) (2) correl = i j σi σj
58
J. Rout et al.
Dissimilarity: Dissimilarity is a feature that evaluates the distance between two items inside the ROI to determine how different they are. The grey level mean difference in the image’s distribution is what this function measures. When the value is bigger, it indicates that the neighboring pixels have a broader range of intensity values to choose from. The characteristic of a picture known as its dissimilarity “Dissimilarity” is defined by the equation as follows: |i − j|p(i, j) (3) Dissimilarity = i
j
Energy: When calculating energy, the amount of times a certain set of pixels appears is taken into account. It quantifies how chaotic a picture is. The energy value is quite high for pixels that are strongly connected. The following equation defines the “Energy” attribute of an image: Energy = p(i, j)2 (4) i,j
Homogeneity: Inverse distinct movement is the term used to describe homogeneity. With bigger values for lower grey tone differences between paired objects, it evaluates picture homogeneity. With constant energy, homogeneity diminishes as contrast rises. The following equation defines the “Homog” homogeneity characteristic of an image: Homog =
1
i
1 + |i − j|2
j
p(i, j)
(5)
3.2 Performance Metrics Algorithms are evaluated using a variety of performance criteria such as accuracy, sensitivity, precision, and F1-score from the confusion matrix. Accuracy: Rate of accurate identification as a percentage of all subjects. Accuracy =
TP + TN TP + TN + FP + FN
(6)
Sensitivity: Recall measures how many correct labels a computer is able to assign. Sensitivity =
TP TP + FN
(7)
Precision: One way to measure an outlook’s reliability is by tallying up the number of correct predictions made thus far. This idea also goes by the name “predictive value”. Pr ecision =
TP TP + FP
(8)
F1-Score: Taking into account both precision and recall, the F1-score is a useful measure. F1 − score = 2 ∗
Pr ecisison ∗ Recall Pr ecision + Recall
(9)
GLCM Based Feature Extraction and Medical X-ray Image Classification
59
Specificity: In this case, the algorithm made the correct determination that the negative should be labeled as specificity. Specificity =
TN TN + FP
(10)
where, TP = True Positive TN = True Negative FP = False Positive FN = False Negative
4 Results Measurements of performance are used to evaluate Ensemble learning’s classification method against others, such Logistic regression, Gaussian Naive Bayes, as well as Random Forest, to prove the algorithm’s efficacy. This comparison is done in order to demonstrate the algorithm’s effectiveness (Fig. 3).
Fig. 3. Accuracy analysis of four algorithms
The accuracy analysis of four algorithms is shown in the above figure, with the suggested and existing algorithms on the x-axis as well as the accuracy value on the yaxis. When compared to other methods, ensemble learning with feature selection gives results that are effective (the accuracy value of ensemble learning is 97) (Fig. 4).
60
J. Rout et al.
Fig. 4. Precision analysis of four algorithms
The precision analysis of four algorithms is shown in the figure above, with the x-axis representing the proposed and current method and the y-axis representing the precision value. The accuracy value of random forest is 90, the precision value of Gaussian naive bayes is 91, and the precision value of logistic regression is 92. When compared to other algorithms, ensemble learning offers excellent results (Fig. 5).
Fig. 5. The F1 score analysis of four algorithms
The above figure represents the F1 score analysis of four algorithms with the x-axis being the proposed and existing algorithms and the y-axis being the value of the F1 score. The F1-score value of ensemble learning is 97, It can be concluded that Ensemble learning with feature selection produces effective results when compared with other algorithms (Fig. 6).
GLCM Based Feature Extraction and Medical X-ray Image Classification
61
Fig. 6. Specificity analysis of four algorithms
From the above figure it is observable that the specificity analysis of four algorithms with the x-axis being the proposed and existing algorithm and the y-axis being the specificity value. Random forest specificity value is 91, Gaussian naïve bayes value is 94 and logistic regression specificity value is 91. Ensemble learning produces effective result compared to other algorithms (Fig. 7).
Fig. 7. Accuracy, Precision, F1-score, specificity performance parameters
Accuracy, Precision, F1-score, specificity performance parameters (x-axis), parameter values (y-axis), and analyses for ensemble learning are shown above. Ensemble learning’s accuracy is 97, its precision is 95, its F-score is 97, and its specificity is 96.
62
J. Rout et al.
5 Conclusion In the current investigation, Features are extracted from medical X-ray pictures using an ensemble learning model and then classified. Image features are extracted by using a feature extraction approach known as GLCM, we were able to differentiate between normal and diseased images (Gray level co-occurrence matrix). In this investigation, we make use of the following GLCM features: contrast, correlation, dissimilarity, energy, and homogeneity. After extracting the features of an image, SMOTE analysis is carried out to balance the classes. After performing SMOTE analysis, proposed model performance is evaluated by using performance metrics. Performance measurements are used to compare the Ensemble learning classification algorithm against other algorithms like Logistic regression, Gaussian Naive Bayes, as well as Random Forest algorithms in order to demonstrate its effectiveness. Ensemble learning’s accuracy is 97, its precision is 95, its F-score is 97, and its specificity is 96.When compared to existing algorithms, the proposed ensemble model was proven to have a 97% accuracy rate in distinguishing between healthy and sick photos.
References 1. Usha, R., Perumal, K.: SVM classification of brain images from MRI scans using morphological transformation and GLCM texture features. Int. J. Comput. Syst. Eng. 5(1), 1 (2018). https://doi.org/10.1504/ijcsyse.2018.10011250 2. Saputra, R.A., Suharyanto, Wasiyanti, S., Saefudin, D.F., Supriyatna, A., Wibowo, A.: Rice leaf disease image classifications using KNN based on GLCM feature extraction. J. Phys. Conf. Ser. 1641(1) (2020). https://doi.org/10.1088/1742-6596/1641/1/012080 3. Aslan, N., Dogan, S., Koca, G.O.: Classification of chest X-ray COVID-19 images using the local binary pattern feature extraction method chest X-ray COVID-19 Görüntülerinin Yerel ˙Ikili model Özellik Çıkarımı Yöntemi Kullanılarak Sınıflandırılması 17(2), 299–308 (2022) 4. Reshi, A.A., et al.: An efficient CNN model for COVID-19 disease detection based on X-ray image classification. Complexity 2021 (2021). https://doi.org/10.1155/2021/6621607 5. Han, Y., Holste, G., Ding, Y., Tewfik, A., Peng, Y., Wang, Z.: Radiomics-guided global-local transformer for weakly supervised pathology localization in chest X-rays, vol. XX, no. Xx, pp. 1–11 (2022). http://arxiv.org/abs/2207.04394 6. Xia, C., et al.: SVM-based bone tumor detection by using the texture features of X-ray image. In: Proceedings of 2018 6th IEEE International Conference on Network Infrastructure and Digital Content, IC-NIDC 2018, vol. 6, no. 62, pp. 130–134 (2018). https://doi.org/10.1109/ ICNIDC.2018.8525806 7. Xing, Y., et al.: Adversarial pulmonary pathology translation for pairwise chest X-ray data augmentation. In: Shen, D., et al. (eds.) MICCAI 2019. LNCS, vol. 11769, pp. 757–765. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-32226-7_84 8. Mall, P.K., Singh, P.K., Yadav, D.: GLCM based feature extraction and medical X-RAY image classification using machine learning techniques. In: 2019 IEEE Conference on Information and Communication Technology, CICT 2019, pp. 1–6 (2019). https://doi.org/10.1109/CIC T48419.2019.9066263 9. Zhang, Q., et al.: A GPU-based residual network for medical image classification in smart medicine. Inf. Sci. 536, 91–100 (2020). https://doi.org/10.1016/j.ins.2020.05.013 10. Zare, M.R., Jehangir, B., Seng, W.C., Mueen, A.: Automatic classification of medical X-ray images. Malays. J. Comput. Sci. 26(1), 9–22 (2013)
GLCM Based Feature Extraction and Medical X-ray Image Classification
63
11. Zeelan Basha, C.M.A.K., Maruthi Padmaja, T., Balaji, G.N.: Automatic X-ray image classification system. Smart Innov. Syst. Technol. 78, 43–52 (2018). https://doi.org/10.1007/978981-10-5547-85 12. Ankita, R., Kumari, C.U., Mehdi, M.J., Tejashwini, N., Pavani, T.: Lung cancer image-feature extraction and classification using GLCM and SVM classifier. Int. J. Innov. Technol. Explor. Eng. 8(11), 2211–2215 (2019). https://doi.org/10.35940/ijitee.K2044.0981119 13. Raju Ahmed, M., Yasmin, J., Wakholi, C., Mukasa, P., Cho, B.K.: Classification of pepper seed quality based on internal structure using X-ray CT imaging. Comput. Electron. Agric. 179, 105839 (2020). https://doi.org/10.1016/j.compag.2020.105839 14. Veena Divya, K., Jatti, A., Joshi, R., Deepu Krishna, S.: Characterization of dental pathologies using digital panoramic X-ray images based on texture analysis. In: Proceedings of Annual International Conference of the IEEE Engineering in Medicine and Biology SocietyEMBS, pp. 592–595 (2017). https://doi.org/10.1109/EMBC.2017.8036894 15. Sumari, P., Syed, S.J., Abualigah, L.: A novel deep learning pipeline architecture based on CNN to detect Covid-19 in chest X-ray images. Turk. J. Comput. Math. Educ. 12(6), 2001–2011 (2021). https://doi.org/10.17762/turcomat.v12i6.4804
Multi-filter Wrapper Enhanced Machine Learning Model for Cancer Diagnosis Bibhuprasad Sahu1(B)
and Sujata Dash2
1 Department of Computer Science and Information Technology, Maharaja Sriram Chandra
Bhanja Deo University, Baripada, Odisha, India [email protected] 2 Department of Computer Application, Maharaja Sriram Chandra Bhanja Deo University, Baripada, Odisha, India
Abstract. The classification accuracy of the high dimensional dataset degrades due to the redundant and irrelevant features. Feature selection (FS) is used to reduce the dimensionality of the dataset by removing the noisy features. Each filter has its statistical approach. So the feature selected by a single filter may ignore the important one. We have presented a multifilter (MF) wrapper hybrid model. The advantage of using the MF method is to select the important feature by one filter which one may ignore by the other. Here, we have used an aggregator approach to combine the most efficacious features among the four individual filter methods (information gain (IG), chi-square (Chi-sq), minimum redundancy maximum relevance (mRMR), and relief). The accuracy assessment is carried out in a multiple filter wrapper (Jaya-SVM, GA-SVM, PSO-SVM, and FA-SVM). The evaluation and prediction of the subset of features are carried out with four classifiers with excellent performance, such as the support vector machine (SVM), Naive Bayes (NB), decision tree (DT), and linear discriminant analysis (LDA) were tested respectively. Four (breast cancer, leukemia, ovarian, and central nervous system (CNS)) cancer datasets are used to implement the model. The performance of the MF wrapper is excellent in comparison to a single filter. According to the findings of this study, the proposed hybrid approach is a more efficient and trustworthy feature selection technique for selecting highly discriminative features. Keywords: Filter · Multifilter (MF) · Wrapper · SVM Classifier
1 Introduction Rapid technological improvements in various domains of life generate data volumes at an unprecedented rate. This may appear to be beneficial to the decision-making process, but it is not when it comes to data dimensions. In microarray data analysis, for example, each sample contains the measurement of tens of thousands of variables, but issues arise in terms of the dimension of the dataset. In the microarray dataset, each sample contains a thousand featured genes. Nevertheless, existing machine learning methods are not designed to cope with huge data sets because a rising ratio of variables © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2023 Published by Springer Nature Switzerland AG 2023. All Rights Reserved S. Nandan Mohanty et al. (Eds.): ICISML 2022, LNICST 470, pp. 64–78, 2023. https://doi.org/10.1007/978-3-031-35078-8_7
Multi-filter Wrapper Enhanced Machine Learning Model
65
to sample size has a severe influence on the capacity to develop models with scientific validity. This is called the “curse of dimensionality”. The microarray datasets contain a high number of features that enhance the noise and directly impact the accuracy of the machine learning algorithm. Such issues can be addressed by feature selection [18]. It decreases data dimensionality by eliminating features that are irrelevant or redundant [20, 21]. Wrappers, filters, and embedded methods are the basic supervised learning models used for feature selection. The filter identifies the features in the preprocessing step and works independently without giving importance to the classification algorithm [19]. But the wrapper and embedded approaches use machine learning algorithms for feature evaluation. Because of their easy ranking procedures, filter approaches such as IG, GR, Chi-square, and Relief-F (RF) have been recommended as the most eminent and convenient filter algorithms for addressing high-dimensional data in recent years. To remove irrelevant features, the filter method uses a statistical ranking score evaluation and a set of threshold values. One ranked feature higher than the threshold value is selected as the significant one and the rest are excluded for classification. In some cases, the performance of the SVM degrades drastically when the number of selected features is either too big or too small. The main reason behind this imbalance is that each filter algorithm always focuses on the dependencies among the features, so it represents poor performance to generate classifier performance. To enhance the performance of the machine learning model, we have used the ensemble MF combined with a robust feature aggregation technique with the wrapper for the selection of optimal features from the microarray datasets.
2 Literature Survey Chyh-Ming Lai et al. [1] presented a VIKOR method that adopted MF (MF) feature subset selection followed by simplified swarm optimization as a wrapper to identify the optimal feature subset. The proposed model was evaluated using the most preferred SVM classifier and achieved 100% accuracy in 15 datasets out of 17 taken into consideration for the experimental study. A multi-stage MF with a wrapper Harris Hawks optimization (HHO) model is proposed by Ali Dabba et al. [2]. After preprocessing with the Min-Max method, followed by five filter approaches (F-score, improved F-score, mutual information (MI), mutual information maximization (MIM), and random forest (RF)), to recognize the top-ranked feature subset from the high-dimensional datasets, followed by wrapper HHO. The performance of the model is evaluated using a support vector machine (SVM) -LOOCV and KNN classifier. Similarly, the author [3] proposed an MF wrapper model to identify optimal feature subsets. In the first stage, three filter methods named CS, IG, and ReliefF are adopted to identify the most informative feature subset by removing redundant features. A nature-inspired algorithm is used to enhance the performance in the second phase of the model. Different classifiers are implemented in parallel with PSO for the evaluation of the proposed model. Two stages of MF supervised gene clustering algorithm is implemented to select the most significant clustered features [4]. In SA-EFS, an aggregation-based ensemble feature selection was proposed. The author used different filter algorithms (chi-square, maximum information coefficient, and XGBoost) [5]. The main focus of the study is the classification of the
66
B. Sahu and S. Dash
microarray cancer dataset by recognizing the informative features. Ab Hamid et al. [6] calculate the correlation between different features by removing the noisy features with the adoption of MF (ensemble) algorithms such as IG, CS, RF, and gain ratio (GR). In the later stage, the evaluation of the proposed one is done using harmonize particle swarm optimization (PSO) and SVM. An ensemble MF feature selection approach named EnRank is proposed by combining the ranked feature subset using T-test, CS, Ridge regression, and LASSO. Later, five classification algorithms are used for the performance evaluation of the model. This model is used for the detection of pulmonary hypertension biomarkers [7]. B. Seijo-Pardo et al. [8] presented a homogeneous and heterogeneous ensemble feature selection approach using a combination method called aggregators, and both methodologies were evaluated using SVM. Sumant, A.S. et al. [9] proposed a multi-stage filter for best feature subset identification. In the first stage of the model, CS and SU filters are hybridized with RF, and in the second stage, the top (5–10)% feature subset is identified. Then the overall model is evaluated using random forest (RF), K-nearest neighbor (KNN), and support vector machine (SVM) classifiers. In [10], the MF concept is introduced in approaches such as univariate (CS, IG, GR, OneR) and multivariate (RF, SVM-AW, SVM-RFE). The high-ranked feature subset is combined using a sampling technique called bootstrap aggregation, followed by the threshold value to identify the final ensemble feature subset. Manosij Ghosh et al. [11] presented a two-phase MF-wrapper model. In phase 1, the top n features are identified using RF, CS, and SU with a union and intersection approach. To find the optimal feature subset wrapper, GA is included in the second stage, followed by NLP, SVM, and KNN. The novel hybrid MF-wrapper Grasshopper optimization algorithm (GOA) is incorporated with SA to handle the slow convergence and exploitation of search done by the GOA. Sharifai, A.G. et al. [12] presented an integrated MF feature selection with wrapper sequential forward selection (SFS) to enhance the performance of the model. In this study, six filter methods are used. Feature rank greater than the threshold value is considered an input for SFS. Moumita Mandal et al. [13] proposed an Ensemble MIRFCS Method, in which three filters called MI, RF, and CS are used to select the best feature from each filter. After a combination of all the features, the efficiency of the model was evaluated using various classifiers such as RNF, KNN, and NB. Uzma et al. [14] proposed a two-stage MF-local search-based Feature Selection (LSFS) wrapper for optimal feature selection followed by SVM, KNN, and NB for evaluation of the proposed model. Like [13], Namrata Singh et al. proposed a multi-stage MF-wrapper model for feature selection. In this study, two variants of filter methods are used, such as subset-based filters (CFS and CONS) and rank-based filters (CS, IG, and RF). Ensemble feature aggregation is carried out after removing duplicate features and then fed to the wrapper-based sequential forward selection algorithm for optimal feature subset selection. SVM with RBF, Random Forest, and KNN classifiers are used for the evaluation of the model [15]. Decorate: a Meta-learning ensemble technique was proposed to achieve better accuracy from the high-dimensional dataset by minimizing the misclassification error [17].
Multi-filter Wrapper Enhanced Machine Learning Model
67
3 MF Aggregator Model Saeys et al. first presented an ensemble feature selector-based feature selection. The key point of adopting an ensemble MF aggregator mode is to prepare a diverse set of feature selections [16]. It is classified into two types, such as homogeneous and heterogeneous ensembles. In the homogeneous, the same base learner is used for a different sample of data, whereas in the heterogeneous, different base learners are used. An aggregator model aggregates the feature weights or ranks. Various aggregator models are used by various researchers, such as weighted mean aggregation (the feature selection is done based on the features having the highest weight mean), and complete linear aggregation. This method employs the complete ranking of all features, followed by the ranks from all ranking lists that are summed for each feature. The best features have the lowest sum of ranks. The robust rank aggregate method finds features that consistently rank higher than expected under the null hypothesis of uncorrelated inputs and gives each feature a significance score. Feature occurrence frequency (feature selection is performed by determining the number of occurrences of every feature across all lists and ranking them based on their frequency of occurrence). And in classification accuracy-based aggregation (feature selection is accomplished by counting the number of occurrences of each feature across all lists and ranking them according to their frequency of occurrence). We used heterogeneous ensembles in this study because the goal of heterogeneous ensembles is to use the best parts of different algorithms to get strong subsets of features. In this study, we have combined four filters named IG, Chi-square, mRMR, and ReliefF to obtain an initial gene subset that can be improved in a subsequent stage by a wrapper approach along with a classifier. 3.1 Algorithm: Aggregator Method
Data: A – No. of filter method Data: B – Threshold (No. of feature to be selected) Result: Combined feature subset of all filters. 1. For each (a) from 1 to A do 2. Calculate the rank (R a ) using filter method (a) 3. End 4. C= Select the ranking (R a) with the ranking combination method. 5. Ct= Identify (T) attributes or features. 6. Input the feature subset to the wrapper.
3.2 Study of Filters In this study, we have used three different filter techniques. The brief study is presented below.
68
B. Sahu and S. Dash
3.2.1 Chi-Square Independence between two features is calculated using the CS (χ2) test. It is one of the statistical tests most preferred if the occurrence of the feature is independent of the class value of the dataset. The (χ2) value of the feature (f) is mentioned in Eq. (1). χ2 =
k c (Yab − Va )2
Va
a=1 b=1
(1)
Let Y presents the number of examples. Yab presents the sample numbers in Ba class with bth interval. The expected frequency is presented as Yab . Here, k and c present the interval count and class count of the dataset, respectively. If proba is the probability occurs for the event j, then the expected value can be evaluated using the formula, Va = T* proba . Here (T) is the total number of events. So the lower chi-square value denotes more dependence between the features. 3.2.2 ReliefF Compared to Relief, ReliefF is the most robust and suitable filter model to deal with incomplete and noisy data. This filtering approach randomly selects an instance Ri and starts searching for k of its neighbors from the same class (called nearest hits Hj ) and a different class called (nearest misses Mj (C)). The weight W [A] for all of its attributes A is then updated depending on Ri value with corresponding Hj and Mj . A Diff function is used to deal with incomplete data and missing value attributes are handled probabilistically. The final equation of ReliefF is presented in Eq. (2). W[A] = W[A] −
k j=1
diff A, Ri , Hj /(m.k) +
C =class(Ri )
k P(C) diff(A, Ri , Mj (C)) /m.k j=1 1 − P(class(Ri ))
(2)
3.2.3 mRMR The MRMR identifies an optimal gene subset with maximal relevance to the target class and minimal redundancy in a gene set. Relevance and redundancy were evaluated using mutual information (I). Let G be the gene subset of a given dataset with class level (c). Let X and Y be the genes in the subset. Then the relevance (Vx ) and redundancy (Wx ) can be evaluated using Eqs. (3) and (4). The mutual information quotient (MIQ) is calculated using Eq. (5). mRMR is used to rank all the features according to their ranks. Vx = I(x, c) Wx =
(3)
1 I(x, y) |S|
(4)
y∈S
maxc MIQx = maxc I(x,c)/ x∈ϕ
x∈ϕ
1 I(x,y) |S| y∈S
(5)
Multi-filter Wrapper Enhanced Machine Learning Model
69
3.2.4 Information Gain (IG) The evaluation of the features is done based on the information gained by considering one feature each time. Information entropy can be used as a metric for feature ranking. Equation (6) represents the entropy of a class feature. H(X) = −
P
(X) log2 P(X)
(6)
Here, P(X) presents the marginal probability distribution for the random variable X. The value of IG for the attribute feature X is then given by Eq. (7). IG(X/Y) = H(X) − H(X/Y)
(7)
where H(X) presents the entropy of dataset X, which quantifies the degree of uncertainty in predicting the value of a random variable. And H(X/Y) presents conditional entropy (i.e., uncertainty based on the known variable Y). Each feature’s order is determined by the IG value, and high-ranking genes are chosen as input features.
4 MF Wrapper: The Proposed Model 4.1 Algorithm: Pseudo Code of Proposed Rank Aggregation Enhanced MF Wrapper Model Input: Training dataset D = {y1 , y2 …yn−1 , xn}, α, filter algorithm set = {IG, mRMR, Chisq, ReliefF}, Classifier = {SVM, DT, NB, LDA} Output: Classify the result of D.
// Stage1: Aggregator Approach 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13.
Procedure aggregation of MFs {} Aggregation F-score { } Number of filters used for aggregation a Ct Number of genes selected by each filter. For filteri 1, a do [score] calculate (filteri , a) [Rank] Rank( score, Ct) {Aggregation Score (Rank, a) Aggregation = Aggregation End for F-score sort[Aggregation] Return F-score End procedure
70
B. Sahu and S. Dash
// Stage2: Wrapper Model 14. Input the feature subset to Wrapper= {Jaya -SVM, PSO-SVM, GA-SVM, PSO - SVM } 15. Obtain the best optimal features.
// Stage3: Classification 16. 17. 18. 19. 20.
Do classify by classifiers, Classifiers = {SVM, DT, NB, LDA} For each classifier, classifier i in the classifier do optimal features from the Foptimal best Select the Learn the classifier i based on dataset D. Return classification result.
Fig. 1. Conceptual framework of the MF wrapper model
In this study, the MF wrapper model is proposed to handle high-dimensional datasets. The idea behind this study is to minimize the search time and the time complexity to extract the feature subset to enhance the accuracy with a lower number of features. In other words, preprocessing of the dataset is carried out for fitness valuation using
Multi-filter Wrapper Enhanced Machine Learning Model
71
statistical measures and feature selection using filter algorithms. The proposed model combines the filter and the wrapper into a hybrid model. It consists of two stages (Fig. 1). • In the first stage, i.e., the filter stage, we used four distinguished filters named IG, Chi-square, mRMR, and ReliefF. We have used an aggregation method to prepare a feature subset that can be considered as input for the wrapper. The main idea behind the implementation of a multi-filter instead of a single filter is that the implementation of a single filter may generate an imbalanced feature subset as it just ignores the impact of the feature subset (selected). And each filter has its statistical approach, so the feature selected in a single filter may ignore the important one. So to avoid this gap, we used MF in the first stage to recognize the important features from the high-dimensional dataset. • In stage two, i.e., the wrapper phase, the final selected feature subset from stage 1 is considered as an input for the wrapper. In this study, we used four different metaheuristic algorithms. For unbiased comparisons of the performance of the proposed model, three-parameter-based (GA, PSO, and FA) and one non-parameter-based (Jaya) metaheuristic algorithm were used. 4.2 Outline of Classifiers In this study, four efficient classifiers were used, namely SVM, NB, DT, and LDA. Different classifiers were carefully selected because their performance varies depending on the type of dataset. And the result is not similar when the dataset varies. To involve the entire dataset, we have preferred 10CV on training and test datasets. 4.3 Experimental Environment Setup Python with a 2.2 Hz Core i7 CPU with 8 GB RAM was used to carry out the overall execution of the proposed hybrid machine-learning model. It is an open-source programming language and contains powerful libraries. To design the experimental study, we used four microarray datasets containing diversity in the number of samples, gene numbers, and classes. After preprocessing (by converting real and categorical values to numeric data, the datasets are prepared for the task), each dataset is randomly divided into two batches (80% as a training set and 20% as a testing set) to estimate the performance based on the confusion matrix. The multi-filter algorithms are parameterized in terms of counts of features in the early stages of our configured model to draw out those amounts of significant features from various features in the datasets. The number of features for each dataset is user-specified, and we have personalized the number for the convenience of our model. The datasets were obtained from the UCI repository. The details of the dataset descriptions are presented in Table 1.
5 Results and Discussions To evaluate the efficiency of the proposed model, different performance metrics such as precision, recall, F-score, confusion matrix, and AUC/ROC graph are used. The metrics [8, 9] are defined as follows Accuracy (Acc) =
True+ve + True−ve True+ve + True−ve + False+ve + False −ve
(8)
72
B. Sahu and S. Dash
F - score =
True+ve +
True+ve 1 2 (False+ve
(9)
+ False−ve )
where, True+ve specifies % of correctly classified (positive examples), False+ve means % of incorrectly classified as positive (but incorrect examples), True−ve means % of correctly classified (negative examples) and False−ve % of means incorrectly classified (positive examples). Table 1. Dataset Study Datasets
Genes #
Classes #
Breast
24481
02 (relapse-34, Non relapse-44)
78
03 (ALL-47, AML-25)
72
Leukaemia3
7129
CNS
7129
Ovarian
Samples #
02 (Survivors-21, Failures-39)
24482
60
02 (Cancer-162, normal -91)
253
In the experiment, we evaluated the performance of the proposed model with four different datasets. From the resulting study in Table 2, we found the average accuracy for breast cancer and ovarian dataset are 95% and 96% respectively, but in the case of leukemia it is 97%. The proposed model performed best in the case of the CNS dataset with 98%. While focusing on the performance of the individual dataset, we found the wrapper FA performs with 100% accuracy, whereas wrapper Jaya and GA achieved an average accuracy of 93% and PSO with 95% respectively. Similarly, in the case of the leukemia dataset wrapper, FA performs with high accuracy with 99%. Other methods such as wrapper Jaya, GA, and PSO perform with an accuracy of 96%, 95%, and 97% respectively. In the case of the CNS dataset wrapper, Jaya and PSO achieve an accuracy of 99%, but wrapper GA and FA perform with 96% and 98% respectively. Wrapper PSO and FA perform with an accuracy of 98% in the case of the ovarian dataset and Wrapper Jaya and GA achieve accuracy of 92% and 97% respectively. In the case of the ovarian dataset, the MF-Jaya with NB classifier performs worst with an accuracy of 76%. Table 2. Performance of MF Wrapper Approach
Dataset
Classifier
Precision
Recall
F-score
Accuracy
MF-Jaya
BC
SVM
0.98
0.98
0.98
0.91
DT
0.98
0.94
0.96
0.93
NB
0.95
0.93
0.94
0.94
LDA
0.97
0.92
0.94
0.95 (continued)
Multi-filter Wrapper Enhanced Machine Learning Model
73
Table 2. (continued) MF-GA
MF-PSO
MF-FA
SVM
0.97
0.97
0.98
0.94
DT
0.94
0.94
0.95
0.91
NB
0.97
0.97
0.96
0.94
LDA
0.98
0.97
1
0.96
SVM
0.96
0.97
0.9
0.92
DT
0.96
0.91
0.9
0.95
NB
0.94
0.93
0.98
0.97
LDA
0.94
0.95
0.94
0.96
SVM
1
0.94
0.96
1
DT
1
0.97
0.96
1
NB
1
0.94
0.98
1
LDA
1
0.97
0.98
1
Approach
Dataset
Classifier
Precision
Recall
F-score
Accuracy
MF-Jaya
Leukemia
SVM
0.9
1
0.94
0.98
DT
1
1
1
0.94
MF-GA
MF-PSO
MF-FA
NB
0.9
1
0.94
0.97
LDA
0.9
1
0.94
0.98
SVM
0.97
0.94
0.96
0.97
DT
0.96
0.91
0.93
0.96
NB
0.95
0.89
0.96
0.92
LDA
0.96
0.97
0.97
0.98
SVM
0.97
0.98
0.97
0.98
DT
0.97
0.98
0.97
0.99
NB
0.98
0.97
1
0.96
LDA
0.94
0.92
0.93
0.96
SVM
0.98
0.99
0.98
0.99
DT
1
0.97
0.96
1
NB
1
0.98
0.95
1
LDA
0.96
0.98
0.96
1
Approach
Dataset
Classifier
Precision
Recall
F-score
Accuracy
MF-Jaya
CNS
SVM
0.9
0.81
0.85
1
DT
0.9
0.81
0.85
1
NB
1
0.9
0.95
1
LDA
0.9
0.9
0.9
0.96 (continued)
74
B. Sahu and S. Dash Table 2. (continued)
MF-GA
MF-PSO
MF-FA
SVM
0.98
0.97
0.97
1
DT
0.89
0.91
0.92
0.94
NB
0.96
0.89
0.91
0.92
LDA
0.97
0.97
0.98
1
SVM
0.97
0.98
0.98
1
DT
0.98
0.98
0.98
1
NB
0.94
0.97
0.98
1
LDA
0.98
0.97
0.98
0.99
SVM
1
1
1
1
DT
0.92
0.94
0.94
0.96
NB
0.89
0.91
0.98
0.99
LDA
0.98
0.98
0.97
1
Approach
Dataset
Classifier
Precision
Recall
F-score
Accuracy
MF-Jaya
Ovarian
SVM
1
1
1
0.99
DT
1
1
1
0.98
MF-GA
MF-PSO
MF-FA
NB
1
1
1
0.76
LDA
1
1
1
0.97
SVM
0.98
0.97
0.98
1
DT
0.98
0.97
0.98
1
NB
0.96
0.94
0.96
0.97
LDA
0.91
0.89
0.91
0.94
SVM
0.96
0.96
0.98
1
DT
0.98
0.97
0.97
0.99
NB
0.92
0.89
0.87
0.93
LDA
0.99
0.98
0.98
1
SVM
0.97
0.98
0.98
1
DT
0.98
0.99
1
1
NB
0.98
1
1
1
LDA
0.92
0.93
0.89
0.94
From the resulting study, it is noteworthy to explain that the MF wrapper performs better than its other counterparts. Average performance accuracy ranges from 92% to 100%. Similarly, the recall and F-score vary between 81% and 100%. The average accuracy achieved by the MF wrapper varies within a range of 95% and 98%. Due to the page limit, it is not possible to present the complete result analysis of each model. We presented the ROC graph of the CNS and breast cancer dataset with the MF wrapper
Multi-filter Wrapper Enhanced Machine Learning Model
75
model. The performance of the hybrid model is compared with the existing models. The comparative study of our model with existing models is presented in Table 3. The performance of our model is mentioned in bold. From Table 3, it is clear that our proposed model performs better except for the ovarian dataset due to overfitting (Fig. 2).
Fig. 2. ROC-AUC curve of MF-Wrapper with different datasets
Table 3. Performance comparisons proposed vs State-of-art Models Dataset
Method
Filter
Classifier
Accuracy
Leukemia
MF-GE
MF
C4.5
84.51
Leukemia
EFHFS
MF
NB
69.3
Leukemia
AC-MOFOA
IG
KELM
96.78
Leukemia
AC-MOFOA
Chi-Square
KELM
96.4
Leukemia
AC-MOFOA
ReliefF
KELM
95.82 (continued)
76
B. Sahu and S. Dash Table 3. (continued)
Dataset
Method
Filter
Classifier
Accuracy
Leukemia Leukemia
AC-MOFOA
mRMR
KELM
96.26
Ensemble PSO
MF
Ensemble
95.3
Leukemia
Wrapper JAYA
MF
All
96(AVG)
Leukemia
Wrapper GA
MF
All
95(AVG)
Leukemia
Wrapper PSO
MF
All
97(AVG)
Leukemia
Wrapper FA
MF
All
99(AVG)
Breast cancer
Ensemble
MF
SVM
72.91
Breast cancer
Ensemble PSO
MF
SVM
96.15
Breast cancer
Wrapper JAYA
MF
All
93(AVG)
Breast cancer
Wrapper GA
MF
All
93(AVG)
Breast cancer
Wrapper PSO
MF
All
95(AVG)
Breast cancer
Wrapper FA
MF
Al
100(AVG)
Breast cancer
AC-MOFOA
IG
KELM
75.11
Breast cancer
AC-MOFOA
Chi-Square
KELM
69.31
Breast cancer
AC-MOFOA
ReliefF
KELM
76.23
Breast cancer
AC-MOFOA
mRMR
KELM
77.29
Breast cancer
Ensemble PSO
MF
Ensemble
64.98
CNS
Wrapper JAYA
MF
All
99(AVG)
CNS
Wrapper GA
MF
All
96(AVG)
CNS
Wrapper PSO
MF
All
99(AVG)
CNS
Wrapper FA
MF
All
98(AVG)
CNS
Ensemble PSO
MF
Ensemble
66.6
Ovarian
Wrapper JAYA
MF
SVM
92(AVG)
Ovarian
Wrapper GA
MF
DT
97(AVG)
Ovarian
Wrapper PSO
MF
NB
98(AVG)
Ovarian
Wrapper FA
MF
LDA
98(AVG)
Ovarian
Ensemble PSO
MF
Ensemble
99.6
6 Conclusions In this study, we have presented an MF wrapper machine learning model for gene selection and classification of high-dimensional microarray datasets. The proposed model begins with an aggregator model, with these high-ranked features selected and combined to prepare one dataset. The outcome of the MF aggregator model is considered as an input for the wrapper. We evaluated our model with four microarray datasets (breast,
Multi-filter Wrapper Enhanced Machine Learning Model
77
leukemia, CNS, and ovarian cancer) using ten CVs. Then the performance of the proposed model is compared with other state-of-the-art machine learning models for the said dataset. The results show that the multiple filters outperform compared with individual ranking methods for initial gene selection. We can conclude that more filters used more possibilities to ensure that no relevant gene was left out or a potential biomarker in the early stages. Based on different experiments during our framework study, it suggests that choosing MF to improvise the performance of the model to select the best feature from the high-dimensional dataset. The experimental results carried out with four microarray datasets demonstrate that adopting this MF wrapper model is not only able to identify the best-featured genes, but also enhances performance. This proposed model is efficient and effective for high-dimensional datasets as per the result analysis in Table 3. In the future, we will use different metaheuristic approaches to deal with noisy, high-dimensional datasets to get the most accurate results by reducing the number of features.
References 1. Lai, C.-M., Huang, H.-P.: A gene selection algorithm using simplified swarm optimization with MF ensemble technique. Appl. Soft Comput. 100, 106994 (2021) 2. Dabba, A., Tari, A., Meftali, S.: A new multi-objective binary Harris Hawks optimization for gene selection in microarray data. J. Ambient. Intell. Humaniz. Comput. 14(4), 3157–3176 (2021). https://doi.org/10.1007/s12652-021-03441-0 3. Alrefai, N., Ibrahim, O.: Optimized feature selection method using particle swarm intelligence with ensemble learning for cancer classification based on microarray datasets. Neural Comput. Appl. 34(16), 13513–13528 (2022). https://doi.org/10.1007/s00521-022-07147-y 4. Bose, S., Das, C., Banerjee, A., Chattopadhyay, M., Chattopadhyay, S.: An ensemble filtering and supervised clustering based informative gene selection algorithm in microarray gene expression data. In: 2020 4th International Conference on Computational Intelligence and Networks (CINE), pp. 1–7. IEEE (2020) 5. Wang, J., Jing, X., Zhao, C., Peng, Y., Wang, H.: An ensemble feature selection method for high-dimensional data based on sort aggregation. Syst. Sci. Control Eng. 7(2), 32–39 (2019) 6. Ab Hamid, T.M.T., Sallehuddin, R., Yunos, Z.M., Ali, A.: Ensemble based filter feature selection with harmonize particle swarm optimization and support vector machine for optimal cancer classification. Mach. Learn. Appl. 5, 100054 (2021). https://doi.org/10.1016/j.mlwa. 2021.100054 7. Liu, X., Zhang, Y., Chunli, F., Zhang, R., Zhou, F.: EnRank: an ensemble method to detect pulmonary hypertension biomarkers based on feature selection and machine learning models. Front. Genet. 12, 601 (2021) 8. Seijo-Pardo, B., Porto-Díaz, I., Bolón-Canedo, V., Alonso-Betanzos, A.: Ensemble feature selection: homogeneous and heterogeneous approaches. Knowl. Based Syst. 118, 124–139 (2017) 9. Sumant, A.S., Patil, D.: Ensemble feature subset selection: integration of symmetric uncertainty and chi-square techniques with RReliefF. J. Inst. Eng. India: Ser. B 103(13), 831–844 (2021). https://doi.org/10.1007/s40031-021-00684-5 10. Pes, B.: Ensemble feature selection for high-dimensional data: a stability analysis across multiple domains. Neural Comput. Appl. 32(10), 5951–5973 (2019). https://doi.org/10.1007/ s00521-019-04082-3
78
B. Sahu and S. Dash
11. Ghosh, M., Adhikary, S., Ghosh, K.K., Sardar, A., Begum, S., Sarkar, R.: Genetic algorithm based cancerous gene identification from microarray data using ensemble of filter methods. Med. Biol. Eng. Compu. 57(1), 159–176 (2018). https://doi.org/10.1007/s11517-018-1874-4 12. Sharifai, A.G., Zainol, Z.B.: Multiple filter-based rankers to guide hybrid grasshopper optimization algorithm and simulated annealing for feature selection with high dimensional multiclass imbalanced datasets. IEEE Access 9, 74127–74142 (2021). https://doi.org/10.1109/ACC ESS.2021.3081366 13. Mandal, M., Ghosh, D., Acharya, S., Saha, N., Sarkar, R.: MIRFCS: an ensemble of filter methods for classification of disease data. In: Das, A.K., Nayak, J., Naik, B., Dutta, S., Pelusi, D. (eds.) Computational Intelligence in Pattern Recognition. AISC, vol. 1349, pp. 205–217. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2543-5_18 14. Uzma, Z.H.: An ensemble filter-based heuristic approach for cancerous gene expression classification. Knowl. Based Syst. 234, 107560 (2021). https://doi.org/10.1016/j.knosys.2021. 107560 15. Singh, N., Singh, P.: A hybrid ensemble-filter wrapper feature selection approach for medical data classification. Chemom. Intell. Lab. Syst. 217, 104396 (2021) 16. Saeys, Y., Abeel, T., Peer, Y.: Robust feature selection using ensemble feature selection techniques. In: Daelemans, W., Goethals, B., Morik, K. (eds.) Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, Berlin, Heidelberg (2008). https://doi.org/10.1007/978-3-540-87481-2_21 17. Dash, S.: A diverse meta learning ensemble technique to handle imbalanced microarray dataset. In: Advances in Nature and Biologically Inspired Computing, pp. 1–13. Springer, Cham (2016) 18. Dash, S., Patra, B., Tripathy, B.K.: A hybrid data mining technique for improving the classification accuracy of microarray data set. Int. J. Inf. Eng. Electron. Bus. 4(2), 43–50 (2012). https://doi.org/10.5815/ijieeb.2012.02.07 19. Sahu, B., Dash, S., Mohanty, S.N., Rout, S.K.: Ensemble comparative study for diagnosis of breast cancer datasets. Int. J. Eng. Technol. 7(4), 281–285 (2018) 20. Dash, S., Patra, B.: Redundant gene selection based on genetic and quick-reduct algorithms. Int. J. Data Mining Intell. Inf. Technol. Appl. 3(2), 1 21. Dash, S., Patra, B.: Feature selection algorithms for classification and clustering in bioinformatics. In: Tripathy, B.K., Acharjya, D.P. (eds.) Global Trends in Intelligent Computing Research and Development, pp. 111–130. IGI Global (2014). https://doi.org/10.4018/978-14666-4936-1.ch005
An Interactive Web Solution for Electronic Health Records Segmentation and Prediction Sudeep Mathew(B) , Mithun Dolthody Jayaprakash, and Rashmi Agarwal REVA Academy for Corporate Excellence, REVA University, Bengaluru, India {sudeep.ba05,mithun.dj,rashmi.agarwal}@reva.edu.in
Abstract. A vast variety of patient data has been collected and monitored through Electronic Health Records (EHR) using various tools in the healthcare. The objective of the paper is to start data acquisition and data understanding and then create a web interface for data exploration and segmentation and classification. In the data modeling phase, the objective is to create machine learning models for segmentation and classification. The first step is data acquisition from the MIMIC-III v1.4 (Clinical database) data mart. In the data understanding phase, the relationship of multiple tables is evaluated. After data wrangling the combined dataset is then used for k-means clustering techniques for obtaining chest heart failure patients clusters. In the following phase, the diagnosis text data is used for data modeling and for that various text features are created and then multiple classification techniques are applied for predicting the occurrences of death and the best model is considered for the model deployment. In the model evaluation phase, it is observed that six clusters were optimal while training the model and it is incorporated into the application for predicting the segments of the patients based on the risk levels. Few machine learning models were trained on patient’s historic diagnosis text data and the logistic regression model indicated 89% of AUC score in test data and is deployed into the application for the prediction. Keywords: Natural Language Processing · EHR · Segmentation · Serious Adverse Event Prediction
1 Introduction Electronic health records (EHRs) contain patient diagnostic records, physician records, and records of hospital departments. For heart diseases, we can receive huge unstructured data from EHR time series. By analyzing and mining, we can identify the links between diagnostic events and ultimately predict the probability of the occurrence of a serious adverse event. The adoption of EHR datasets and the increase of digitized information about patient data revolutionize the emergence of clinical research in oncology research [1]. One of the applications of EHR data is an improvising learning system for clinical research and which helps in various applications of patient selection, dosing, drug target, etc. as discussed [2]. Standardizing electronic health records in the Indian health record system is implemented by [3]. The comprehensive techniques for modeling EHR data are provided by [4]. © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2023 Published by Springer Nature Switzerland AG 2023. All Rights Reserved S. Nandan Mohanty et al. (Eds.): ICISML 2022, LNICST 470, pp. 79–91, 2023. https://doi.org/10.1007/978-3-031-35078-8_8
80
S. Mathew et al.
The web app is developed by aiming to help the medical or clinical team to monitor the safety of the patients during the clinical trials by allowing the users to explore the data through visualization and statistics, clustering, and the probability of the occurrence of SAE. Clustering or segmentation techniques are helpful to find out the underlying hidden association between each data point. This helps the business to decide on each cluster. Detection anomalies or outliers in EHR data was implemented by [5]. The primary objective is to apply unsupervised clustering techniques to EHR data the result indicated that clustering techniques produced high sensitivity and specificity. The occurrence of serious adverse events or SAE is one of the primary concerns that pharmaceutical companies face during the clinical trial. This application intended to solve this problem by providing the probability of the occurrence of SAE by analyzing patient’s diagnosis data. In one of the works, a prediction model was implemented to detect the occurrence of an adverse event such as cardiac arrest by utilizing patient data [6]. The paper primarily focuses on the clinical research industries team which helps to improve overall the patient’s safety concerns and address key issues during the clinical trial.
2 Background 2.1 Application of ML on HER Data One of the works by Ziyi was referring to the challenges and the perspectives of machine learning multimodal in electronic health records. And this work suggests that including structured data is not enough to achieve a good result instead this study seeks to use machine learning and deep learning models on structured and unstructured EHR datasets [7]. The machine learning models in electronic health records could outperform conventional survival models for predicting mortality in coronary disease. These works include multiple machine learning models such as the cox model, and random forest in the 80000 patients EHR dataset and the output indicated that it outperforms conventional models [8]. Adeler proposed a risk prediction model for chronic kidney disease progression using heterogeneous electronic health record data and time series analysis. This work was to combine the time series model and cox proportional model to get a range of risk prediction models and the evaluation of the model is based on discriminatory statistics [9]. Predicting the risk of heart failure with EHR sequential data modeling is developed and used patient’s diagnosis data with the LSTM sequential model used and the evaluation is based on the utility and efficacy of the proposed solution [10]. 2.2 Unsupervised Techniques on HER Data Lutz Has performed an unsupervised machine learning model to detect patient subgroups in electronic health records. This project used agglomerative hierarchical clustering and k-means clustering on patient’s lab and coded datasets. The results indicated that natural grouping is present in the dataset and hierarchical clustering provides higher quality
An Interactive Web Solution for Electronic Health Records
81
clusters than k-means clustering [11]. One of the works implemented by Gabriele is for private hospital workflow optimization through k-means clustering. This work is to optimize work allocation-based staff members, patients, hospitals, and locations to cluster the staff members utilizing the frequency of the facing time [12]. k-means clustering is used for healthcare knowledge discovery is one of the works and the objective is used to discover hidden patterns by applying k-means clustering and a self-organizing map (MAP) [13]. One of the works is an unsupervised machine learning model used for the discovery of latent disease clusters using electronic health records [14]. Prediction of health outcomes for pediatric patients is one of the works implemented and the approach of the project was to implement a Bayesian model and clustering for predicting the risk of type 2 diabetics for children between the ages of 10 and 14 [15]. Another work implemented using an unsupervised LDA approach to cluster patient subgroups into multiple clusters using patient’s health records [16]. 2.3 Text Analytics on HER Data Natural language processing is used in the field of unstructured text data and this work explores the possibilities of applying NLP models in EHR datasets. The author discussed various applications of NLP such as classification models, question answering, phenotyping, knowledge graphs, medical dialogue, etc. [17]. A comparative analysis of text classification approaches in electronic health records indicated that text classification in traditional approaches could exceed the performance of contextual embedding models such as BERT [18]. A work mentioning the application of deep learning models in an electronic health record by developing various deep neural network models. The proposed solution is a deep learning model for predicting the health risk of the patient [19]. On other hand, a work implemented by Mascio suggests different text classification approaches to patient text data [20]. This work focused on various traditional machine learning models and compared them with contextual embedding BERT and identified that traditional models performed well on the text data. Bittar tried to implement suicide risk assessment using text data and the work implemented to predict the tendency to commit suicide by extracting text features from the clinical notes and the data trained SVM model [21]. Various machine learning techniques on electronic health records were discussed above on segmentation as well as text classification. This paper seeks to predictability of death of the patient based on the historic text data as well segmentation of HER data.
3 Methodology In contrast to the above-mentioned methods, we develop a k-means clustering technique to group Congestive Heart Failure (CHF) patients based on risk levels. We also developed a Serious Adverse Event (SAE) prediction, model for predicting the probability of death using the text classification technique. We present the details of our approach in the following session. In Fig. 1 the details of the application are represented. In the following sessions, the details of the application describing.
82
S. Mathew et al.
Fig. 1. Application Interface
3.1 K-means Clustering on CHF Data The input data for the clustering or segmentation technique is combined data of patient demographic, admission, diagnosis, and drug administration. The following flowchart in Fig. 2 depicts the proposed methods for the k-means clustering technique.
Fig. 2. Methodology for K-means Clustering
The input features containing patient’s demographic, admission, prescription, and diagnosis datasets were joined and transferred to Principal Component Analysis (PCA) model for dimensionality reduction, and afterward, the principal components are transferred to the k-means model for the segmentation of the patients. This model helps the business to identify the risk levels of the patients to take a valid decision for each cluster. Data Pre-processing and Feature Engineering. The output of the data wrangling created a single data frame for the data analysis and data modeling. The primary diagnosis of the patients with this CHF disease formed a data frame in the previous data wrangling steps. To perform segmentation, data scaling and missing records removal were the two methods performed before the data modeling phase. Data scaling is the approach to normalizing the range of independent features of the dataset. And finally, all the missing values are removed from the data frame by the missing value removal method in python. Table 1 lists all the features that were generated.
An Interactive Web Solution for Electronic Health Records
83
Table 1. Features of Segmentation No Name
Description
1
Count of diagnosis
Aggregated diagnosis count for each patient
2
Drug administrated days Aggregated total days of drugs given for each patient
3
No of drugs
Total number of drugs given to each patient
4
Age Group
Grouped patients based on age such as >90 as very old, 60–80 as a senior citizen, >18 as an adult,