Interpretable Cognitive Internet of Things for Healthcare 9783031086366, 9783031086373, 3031086368

Citation preview

Internet of Things

Utku Kose Deepak Gupta Ashish Khanna Joel J. P. C. Rodrigues   Editors

Interpretable Cognitive Internet of Things for Healthcare

Internet of Things Technology, Communications and Computing

Series Editors Giancarlo Fortino, Rende (CS), Italy Antonio Liotta, Edinburgh Napier University, School of Computing, Edinburgh, UK

The series Internet of Things - Technologies, Communications and Computing publishes new developments and advances in the various areas of the different facets of the Internet of Things. The intent is to cover technology (smart devices, wireless sensors, systems), communications (networks and protocols) and computing (theory, middleware and applications) of the Internet of Things, as embedded in the fields of engineering, computer science, life sciences, as well as the methodologies behind them. The series contains monographs, lecture notes and edited volumes in the Internet of Things research and development area, spanning the areas of wireless sensor networks, autonomic networking, network protocol, agent-based computing, artificial intelligence, self organizing systems, multi-sensor data fusion, smart objects, and hybrid intelligent systems. Indexing: Internet of Things is covered by Scopus and Ei-Compendex **

Utku Kose • Deepak Gupta • Ashish Khanna • Joel J. P. C. Rodrigues Editors

Interpretable Cognitive Internet of Things for Healthcare

Editors Utku Kose Department of Computer Engineering Suleyman Demirel University Isparta, Turkey Ashish Khanna Department of Computer Science Engineering Maharaja Agrasen Institute of Technology Rohini, Delhi, India

Deepak Gupta Department of Computer Science Engineering Maharaja Agrasen Institute of Technology Rohini, Delhi, India Joel J. P. C. Rodrigues Campus Petronio Portela, Bloco 8 Federal University of Piaui, Centro de Tecnologia Teresina, Piauí, Brazil

ISSN 2199-1073 ISSN 2199-1081 (electronic) Internet of Things ISBN 978-3-031-08636-6 ISBN 978-3-031-08637-3 (eBook) https://doi.org/10.1007/978-3-031-08637-3 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Foreword

In the twenty-first century, Artificial Intelligence has gained an exponential momentum, changing its nature and interpretability at the view of users. In the context of especially Deep Learning, interpretability level of advanced intelligent systems reduced because of the trade-off between accuracy and the number of parameters determining the interpretability level. Although simpler Machine Learning models allow us to interpret their outcomes to evaluate their long-time usage, Deep Learning and hybrid models are black-box versions requiring additional efforts to understand the connection between inputs and outputs. Currently, that’s tried to be solved with efforts by interpretable additions of newly developed models of a trendy research area: Explainable Artificial Intelligence (XAI). On the other hand, normal users have been also under the attack of daily life smart tools since the start of twenty-first century. As a result of rapid advances in terms of hardware components, we have been experiencing the era of communication devices through Internet of Things (IoT) technology. At this point, it is a need to find a common relation between interpretability/XAI level of smart tools and their modeling under the view of IoT. Furthermore, cognitive capabilities to ensure better, human-like (or humancompatible) versions of such smart tools have been widely examined in recent research studies. So, all these research orientations trigger the scientific audience to build up some reference works evaluating the most recent knowledge of applications in the context of some critical fields. This edited book Interpretable Cognitive Internet of Things for Healthcare takes the mentioned research considerations to the field of healthcare, and ensures a timely contribution to the fields of both artificial intelligence and medical. The field of medical has a critical role in terms of humankind and the environmental impact. Furthermore, healthcare applications employ the most important applications types, as targeting well-being of humans under applications such as diagnosis, treatment, and management (of the medical processes). As including IoT-based components and enabling such components to simulate somehow cognitive capabilities, the latest research works often take the known findings more steps away to shape the smart future. As following these aspects, the book employs a total of 10 chapters aiming to report internationally done research works, by considering interpretability, cognitive v

vi

Foreword

solution approaches, and of course use of Artificial Intelligence and IoT-combined set-ups to solve target healthcare problems. I liked the most about the topic coverage as some very recent healthcare issues such as COVID-19, personal requirements, robotic systems, and image-oriented automated solutions are all covered very well in the book. The book also hosts some critical review works informing the readers about both essentials and advanced sides of interpretable cognitive IoT usage for specific healthcare problems. I think the coverage of the book will be a unique for the associated literature, so researchers, professionals, and experts from computer science, electrical engineering, biomedical engineering, robotics, communication engineering, and of course medical will be finding it an essential source to be used along their courses as well as projects. Additionally, the book will be very useful for also degree students, who will be more responsible for safe, interpretable use of cognitively improved IoT healthcare tools of the future. As my final words, I would like to thank the editors Dr. Kose, Dr. Gupta, Dr. Khanna, and Dr. Rodrigues for their timely contribution to the literature. Also, special thanks go to all the authors. I wish all readers an enjoyable reading! Afyon Kocatepe University, Afyonkarahisar, Turkey

Omer Deperlioglu

Preface

As a result of rapid advances in terms of artificial intelligence and communication technologies, it has been easier to develop communicating devices, which are capable of understanding the data and deriving outcomes to act further for next actions within digital ecosystems. Nowadays, there are many remarkable developments affecting common technology areas and our life standards. Especially rise of artificial intelligence with Deep Learning and the latest formations of Internet of Things technology have been changing the way of processing digital data. Thanks to such technologies, we are all moving towards an automated future employing human and intelligent system collaboration for advanced technological solutions. As design of artificial intelligence systems has been widely affected by the human, as taking part in the center of all other natural inspiration sources, the latest research ways found their way already in our cognitive characteristics. That has been critical also for rising more smart tools communication other smart devices like humans. For especially a few recent years, cognitively improved Internet of Things has been used in different fields. As it has a vital role in human life, healthcare is among these fields. In the context of all research efforts regarding cognitive Internet of Things use in healthcare applications, one critical problem is ensuring interpretability in the solution framework. Interpretability has been often discussed because of black-box Deep Learning systems, which are making it difficult to understand the connection among input and output data so that being sure about safety and sustainability of the used smart tool. That becomes too critical within healthcare applications as small errors may cause fatal problems at patient side. Using interpretability is useful for not only medical staff but also managers, developers, and all other stakeholders taking place at different points of the wide healthcare service processes. As following the mentioned current state of the literature, we are happy to provide our edited book volume Interpretable Cognitive Internet of Things for Healthcare. This book is a collection of different research works done by international authors to employ necessary interpretable features inside their cognitive Internet of Things solutions for specific healthcare and medical problems. We believe that the book will be beneficial for researchers from engineering fields and vii

viii

Preface

medical. It also comes with deep reviews, and both theoretical-practical knowledge to inform public or private sector workers and experts to learn about the latest state of artificial intelligence, communication technologies, and healthcare. We also invite BSc., MSc., PhD., and post-doc students to consider this book among their essential reference works. In detail, we have reviewed and included a total of 10 chapters including the following research aspects: Chapter “Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review” takes the concept of interpretable artificial intelligence to the latest version: explainable artificial intelligence (XAI), and provides a review for the use of Internet of Health Things. Chapter “ARIMA and Predicted Geospatial Distribution of COVID-19 in India” includes use of a simple, interpretable ARIMA model to detect geospatial distribution of COVID-19 in the region of India. Chapter “Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in Healthcare IoT Systems” follows another important issue: sensitive data and cybersecurity aspects in the context of Internet of Things. In detail, it introduces a multi-party computation-based system architecture for privacy preserving data analysis in healthcare applications. Chapter “A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos: A COVID-19 Solution” considers the image-related cognitively processing aspects and reports a deep learning solution to monitor social distancing (regarding COVID-19 pandemic) in real-time videos. Chapter “Osteoarthritis Detection and Classification in Knee X-Ray Images Using Particle Swarm Optimization with Deep Neural Network” introduces a remarkable hybrid system usage for a specific disease diagnosis. In this context, it reports the osteoarthritis diagnosis in knee X-ray images, done via particle swarm optimization-supported deep neural network. Chapter “Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques” considers use of interpretable Machine Learning models to predict the VLDL cholesterol value. It builds up specific research flow to show effects of different models for the target prediction issue. Chapter “A Survey of Interpretable Cognitive IoT Devices for Personal Healthcare” ensures another wide review evaluation for interpretable cognitive Internet of Things tools to be used for personal healthcare issues. Chapter “Application of Big Data Analytics and Internet of Medical Things (IoMT) in Healthcare with View of Explainable Artificial Intelligence: A Survey” recalls another hot topic: Big Data. It informs the readers about the latest applications of Big Data analytics in Internet of Medical Things with the view by explainable artificial intelligence. Chapter “An Interpretable Environmental Sensing System with Unmanned Ground Vehicle for First Aid Detection” proposes a robotic system formation by considering interpretability issues. It reports a set-up of environmental sensing system with unmanned ground vehicle for first aid detection.

Preface

ix

Chapter “Impact of Pandemic Over Cognitive IoT for Healthcare Industry: A Market Study” focuses on the market factor, and it comes with a review about the impact of pandemic over cognitive Internet of Things for healthcare industry. As it can be seen, the book employs a wide topic scope in the context of cognitive Internet of Things used for alternative healthcare applications. We did our best to ensure some recent review work and come with chapters based on critical, recent issues. We thank to all respectful authors for their valuable chapter contributions. Dear Prof. Deperlioglu from Afyon Kocatepe University, Turkey gave a kind evaluation for the content of the book. We send our warm thanks to him. As the editors, we are all open for ideas as well as suggestions from all readers around the world. Please turn the pages to get the latest technical knowledge about one of the hottest topics in the context of artificial intelligence and healthcare. Thanks for your interest! Isparta, Turkey Rohini, Delhi, India Rohini, Delhi, India Teresina, Piauí, Brazil

Utku Kose Deepak Gupta Ashish Khanna Joel J. P. C. Rodrigues

Acknowledgement

As the editors, we would like to thank Mary James, Hemalatha Velarasu, and the Springer Team for their valuable efforts and great support on pre-organizing the content and publishing of the book.

xi

Contents

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Subrato Bharati, M. Rubaiyat Hossain Mondal, Prajoy Podder, and Utku Kose ARIMA and Predicted Geospatial Distribution of COVID-19 in India . . . . Prisilla Jayanthi and Iyyanki MuraliKrishna Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in Healthcare IoT Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kevser Sahinbas and Ferhat Ozgur Catak A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos: A Covid-19 Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Usman Ahmad Usmani, Junzo Watada, Jafreezal Jaafar, Izzatdin Abdul Aziz, and Arunava Roy Osteoarthritis Detection and Classification in Knee X-Ray Images Using Particle Swarm Optimization with Deep Neural Network . . N. Hema Rajini and A. Anton Smith

1

25

57

73

91

Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 ˙ Ilhan Uysal and Cafer Çali¸skan A Survey of Interpretable Cognitive IoT Devices for Personal Healthcare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Rav Raushan Kumar Chaudhary and Kakali Chatterjee Application of Big Data Analytics and Internet of Medical Things (IoMT) in Healthcare with View of Explainable Artificial Intelligence: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 Anurag Sinha, Den Whilrex Garcia, Biresh Kumar, and Pallab Banerjee

xiii

xiv

Contents

An Interpretable Environmental Sensing System with Unmanned Ground Vehicle for First Aid Detection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 Ali Topal, Mevlut Ersoy, Tuncay Yigit, and Utku Kose Impact of Pandemic Over Cognitive IoT for Healthcare Industry: A Market Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 Deepshikha Bhargava and Amitabh Bhargava Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review Subrato Bharati, M. Rubaiyat Hossain Mondal, Prajoy Podder, and Utku Kose

1 Introduction “Smart healthcare” is a term that refers to the utilization of technologies, that is, artificial intelligence (AI), Internet of Things (IoT), and cloud computing, in order to allow an efficient, convenient, and customized healthcare system [1, 2]. These technologies enable real-time health monitoring via the use of healthcare apps on smartphones or wearable devices, empowering people to take control of their health. Health data gathered at the user level may also be shared with doctors for further diagnosis [3] and, when combined with AI, can be utilized for health screening, illness early detection, and treatment plan selection [4]. In the healthcare sector, the ethical problem of AI transparency and a lack of confidence in the black-box functioning of AI systems necessitates the development of explainable AI models [5]. The approaches of artificial intelligence used to explain AI models and their predictions are referred to as explainable AI (XAI) methodologies [6]. Globally, healthcare costs are increasing at an alarming rate. The prevalence of chronic disease is increasing, and new costly therapies are being created at a breakneck pace. As a result, healthcare professionals across the globe anticipate a gloomy future for healthcare systems. Healthcare will be enhanced and made more cost-efficient via the use of artificial intelligence (AI), mitigating the effects of these advances. Clinical decision-making assisted systems (CDASs) employ artificial intelligence (AI) to assist physicians in making diagnosis and treatment decisions in

S. Bharati () · M. R. H. Mondal · P. Podder Institute of Information and Communication Technology, Bangladesh University of Engineering and Technology (BUET), Dhaka, Bangladesh e-mail: [email protected]; [email protected] U. Kose Department of Computer Engineering, Suleyman Demirel University, Isparta, Turkey e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_1

1

2

S. Bharati et al.

clinical practice. Rather than relying on a pre-existing knowledge base to match patient characteristics to a CDAS, an AI-based CDAS makes use of AI models trained on real-world patient data. Nonetheless, artificial intelligence (AI) is not a silver bullet. As shown throughout history, technological progress is inextricably linked to new problems and significant challenges. While some of these issues pertain to AI’s technological characteristics, others relate to patient, medical, and legal viewpoints, and it is essential to take a multidisciplinary approach. Today’s deep learning systems are efficient for generating very accurate results, yet their behavior is frequently difficult to decipher because of their extreme opaqueness, if not complete invisibility. Even experienced specialists may struggle to fully understand these known “black-box” models. As these deep learning techniques are widely used, academics and policymakers may question if the accuracy of a particular task trumps other critical aspects in the decision-making process. Since ethical standards must be integrated into AI design and implementation, policy discussions are increasingly focusing on Trustable AI, which includes Responsible AI, Valid AI, and Explainable AI (XAI), as well as Privacy-Preserving AI, the latter of which seeks to answer the fundamental question of how decisions are made (Fig. 1). For example, in the United Kingdom, similar demands came from the House of Lords’ AI Team, which stated that developing understandable AI schemes is a necessary condition for AI to be incorporated as a reliable device for the people. The EU’s High-Level Group on AI has launched further research on the road to XAI (Fig. 1). Some organization is funding a new research initiative aimed at developing better explainable artificial intelligence. Because of the growing use of AI-enabled decision-making tools in clinical trials and other real-life situations where humans are making important choices, these arguments will only intensify in the future. Meanwhile, AI research continues to advance at a steady clip. XAI is a thriving field, with various ongoing research and several novel methods developing that have a significant effect on AI improvement in a variety of ways. Though the word is used inconsistently, it is a group of procedures that inform how an AI system makes decisions and predicts the future. In the future, XAI delves into the rationale behind the decision-making procedure, discusses the system’s advantages and disadvantages, and provides a peek into how the system may behave. By providing understandable descriptions of how AI schemes conduct their research, XAI enables investigators to comprehend the visions generated by their findings. For instance, in Fig. 2, to make the learned model more visible and reliable, an explainable AI module may be included. By introducing an explainable AI, both human experience and generalization error may be incorporated in a validated prediction, while a conventional machine or deep learning model just includes generalization error. In comparison, a learned black-box model without an explainable surrogate module would raise worries among end-users, even if the taught model performs well. In research, the general public, and policy discussions, a number of words are used to describe desirable XAI system features, including the following: • Explainability: It explains how a choice was reached to a broader group of users. • Interpretability: It is a feeling of understanding how AI works.

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

3

Fig. 1 Trustworthy AI such as Responsible AI, Valid AI, Explainable AI (XAI), and PrivacyPreserving AI

Fig. 2 A basic system of XAI for DL or ML

• Justifiability: It shows that one has done the research to back up his or her arguments. • Transparency: It assesses how easily data or models may be accessed. • Contestability: The implication is that users may provide arguments against a choice.

4

S. Bharati et al.

The main focus of this chapter is to introduce literature reviews in Sect. 2, where we have discussed existing reviews or surveys on XAI. Next, Sect. 3 illustrates IoHT for healthcare in XAI and XAI in healthcare for different purposes are described in Sect. 4. Then, conclusions are provided in Sect. 5.

2 Related Works Due to the growing availability of healthcare data and the rapid development of analytical methods, artificial intelligence (AI) is attempting to replicate human brain processes in healthcare, causing a paradigm shift. We take a quick look at where healthcare artificial intelligence (AI) applications are at the moment and where they are headed in the future. The authors of [7–12] provide an up-to-date and thorough review of medical image exploration. In the healthcare community, artificial intelligence (AI) has long been a source of contention [13–15]. To “read” features from a large volume of medical data, sophisticated algorithms can be improved by applying AI, and the information gathered can then be used in clinical practice. AI may incorporate learning and capabilities of self-correcting in order to improve accuracy in response to feedback. An artificial intelligencepowered gadget [16, 17] will aid clinical decision-making by providing current medical information from professional, manual, and journal procedures in order to advise on optimal patient care. Additionally, AI systems can help to moderate the inevitable therapeutic and medical errors that occur in human clinical practice (by making decisions more objective and repeatable) [17–20]. AI systems are capable of managing critical data gathered from a large patient population in order to assist in making real-time inferences for health risk warnings and outcome prediction [21–23]. Clinical settings are being rapidly infiltrated with cutting-edge AI-powered technologies, as AI has recently resurfaced in scientific and popular consciousness. Despite this, artificial intelligence in healthcare has been hailed as one of the most promising future areas. Researchers have proposed and developed a variety of clinical decision-assisted systems since the mid-twentieth century [24, 25]. Since the 1970s, rule-based techniques have accomplished a great deal, including electrocardiogram (ECG) interpretation, disease identification, treatment selection, and providing scientific logic explanations (SLXs). On the other hand, developing rule-based systems is time-consuming and error-prone, as it requires the explicit formulation of decision rules. Higher-order relations between disparate pieces of data provided by disparate experts are also challenging to encode, as well as the configurations’ effectiveness is limited by the breadth of earlier medical knowledge [26–36]. Additionally, incorporating a method that associates probabilistic and deterministic reasoning processes was challenging in order to contract down the suitable psychological context, prescribe treatment, and prioritize medical theories [37, 38]. While the initial generation of AI systems was entirely focused on expert curation of medical information and the creation of rigid decision rules, subsequent AI

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

5

research has included machine learning (ML) methods that can explain complex interconnections [39–45]. This enables the algorithms to deduce patterns from clinical data. The machine learning model learns to generate the accurate output for each new occurrence of a labeled input-output pair. This type of supervised machine learning technique is used to determine the optimal model parameters in order to minimize the discrepancies between their predictions and actual results in training instances, with the goal of generalizing the discovered correlations to situations not comprised in the training dataset. The ML model’s generalizability may then be assessed using the test data. Typical objectives for developing supervised ML models include regression, classifying, and defining the similarity of two instances. Keep in mind that while dealing with a supervised dataset, low-dimensional representations may be more readily identified. Without specifying decision-making rules for individual tasks or accounting for complicated connections between input features, artificial intelligence (AI) applications may be created utilizing machine learning techniques to investigate previously undiscovered data patterns. As a consequence, machine learning has become the main technique for emerging AI technologies. The present renaissance of artificial intelligence is mainly due to the active use of deep learning, which entails training a multi-layer artificial neural network (i.e., a deep neural network) on huge datasets [46]. Traditional neural networks are becoming more complex, with some networks containing up to 100 layers. Additional data, advanced architectural designs, or processing time may be required to achieve higher performance when simulating complex input-output interactions using multi-layer neural networks. As a consequence, contemporary neural networks often have parameter counts of several hundred thousand to several hundred million and need a significant amount of computing power to train effectively. Recent advancements in the architecture of computer processors have made it simpler to harness the computing power required for deep learning. While this is true for unlabeled cases, deep learning algorithms are still extremely “datahungry.” Massive medical database repositories suitable for these algorithms have only recently become available as a result of the establishment of various large-scale research projects [47, 48] and data collection platforms. Due to the rapid advancement of artificial intelligence, our ability to build sophisticated models using collected medical data now allows us to automate diagnosis while also providing a more accurate method to treatment by targeting services and personalizing treatments for optimum efficacy in a timely and dynamic way. The list of possible applications in Fig. 3 is not comprehensive. While artificial intelligence (AI) has enormous potential for improving medical care, many technological hurdles remain. We must make an effort to collect data that represents the patient population since deep learning techniques need a large quantity of highquality training information. As a result, a model trained on data from one hospital may be inapplicable to data from another due to the fact that the data from those hospitals include varying degrees of bias and noise [49]. Consensus diagnostics may substantially enhance the training efficiency of deep learning models when there is a lack of inter-expert agreement on the diagnostic function [50]. Appropriate curation

6

S. Bharati et al.

Fig. 3 Possible application of XAI in healthcare

is essential for the management of diverse data. A true gold standard for assessing a patient’s clinical condition may be reached only by having physicians review their clinical findings on an individual basis, which may be prohibitively costly when applied to a large population. A new gold standard has been suggested lately [51], using diagnostic codes and natural-language processing techniques. In lifeor-death situations, advanced algorithms capable of handling the quirks and noise inherent in various datasets may significantly enhance the safety and efficiency of prediction algorithms. The bulk of recent advancements in neural networks have been focused on well-known tasks that do not need data amalgamation from many sources and modalities. General diagnoses (i.e., the study of signs and symptoms, previous medical history, clinical course, and test results) and treatment planning using deep neural networks are more complex than they seem at first glance. Clinical treatment and diagnosis tasks generally need more attention (e.g., medical history, beliefs, and patient interests) than the normally inadequate tasks that DL excels at (e.g., image identification, translation, voice recognition, and sound synthesis). The ability of transfer learning methods to transform models from massive nonhealthcare datasets into algorithms for the study of multimodality clinical datasets is also unclear. More extensive data annotation and data collecting efforts are required for the development of end-to-end medical AI systems. It is notable that Graph Neural Networks are being used as the primary technique for integrating multimodal capability information [52].

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

7

3 IoHT for Healthcare in XAI With recent advancements in IoHT and 5G technology, it is now likely to send sensory data from a health sensor to get feedback, remote clinician, and return a control signal for actuators in milliseconds. Providers will be able to offer healthcare in real time regardless of their geographic or location distance from one another [53, 54]. To make 5G vertical a reality, software-defined networks, network slicing, and network management for 5G radio access networks will all operate in concert. Recent advances in artificial intelligence (AI) algorithms have depicted dependable and promising results when it comes to identifying hidden potential values in massive quantities of raw IoHT data. At many levels, modern artificial intelligence (AI) algorithms allow the healthcare ecosystem to be automated. The precision of deep learning (DL) algorithms has been shown to be exceptional in healthcare automation [30–32, 34, 40]. Diabetes retinopathy (DR) is classified into five stages: severe DR, moderate DR, none, mild DR, and proliferation-associated DR [55]. Images of the fundus may be used to identify these five kinds of cataracts. Despite the fact that human ophthalmologists can easily distinguish between Mild DR, No DR, and Proliferative DR with the naked eye, even medical experts have difficulty distinguishing between Severe DR and Moderate. Due to the fact that each class has its treatment plan, even a little mistake by human actors may have far-reaching consequences [56]. However, after being taught many fundus images from the five categories mentioned above, deep learning algorithms provide higher results with precise low false positive/negative rates and precise high accuracy. Due to the improved accuracy of deep learning algorithms, inferencing the DR class may be completed in less than a second or minute, as compared to the months or years it may take for a patient to find an appointment, resulting in financial damages and deterioration of eye problems. As a consequence, medical diagnosis based on deep learning is becoming more popular and credible [57]. On the other hand, before deep learning and IoHT-based healthcare applications can be adopted, they must show their viability to healthcare professionals, patients, and society as a whole [58]. Despite the fact that there are many paths to sustainability, this research focuses on three critical areas: safeguarding health-related applications and data, preserving users’ privacy, and being generally acceptable [59]. These applications need specialized computer equipment, such as a graphics processing unit (GPU), in order to train the model on a huge training dataset. As a consequence, conventional deep learning systems need the transfer of private and sensitive health data to a cloud server equipped with a GPU for training. It is then utilized for inferencing by storing the model on the cloud after it has been trained. Test data and specimens are often uploaded to the cloud by clients that need decision-making processes [60]. For instance, despite the fact that this compromises data privacy and security, biometric or health data must be distributed with the supplier of deep learning applications [61]. There are numerous social norms and regulations concerning the collection of health data. It is not unusual for the deep learning model to be trained on images of over 60,000 distinct patients’ eyes, each with an exceptional biometric.

8

S. Bharati et al.

As a consequence, technological developments must be guided by ethical, security, legal, and privacy concerns. If the training dataset is contaminated, deep learning algorithms are vulnerable to prejudice. To what extent should an oncologist or a deep learning algorithm be held responsible if a cancer model is shown to be defective and makes an erroneous positive or negative prediction, such as diagnosing stage-4 cancer with a non-cancer patient and starting treatment as a consequence. If appropriate security and privacy protections are not provided, assailants may attack all three stages of dataset collection, training, and inference, making it undesired by society. One approach for increasing the credibility of health data is to exclude sensitive and private information from the training dataset. This protects the data exchanged with DL systems. As a consequence, AI applications in healthcare will acquire confidence and become more generally adopted [62–65]. Federated learning (FL), which allows for the retention of private data on the client’s edge device, has been gaining popularity in recent years [66, 67]. Rather than that, an early model is provided with users, who then train it on their private data before returning it. Thus, rather than transmitting data, it shares the model with customers, posing very little danger of sensitive information being leaked [68]. During differential privacy, a tiny amount of private data may be revealed, depending on the quantity of noise added [69]. As a consequence, researchers have devised innovative ways for increasing privacy while simultaneously increasing sustainability. One approach is to use encrypted computing for the training process, encrypt the dataset, and then train the encrypted model. Fully homomorphic encryption, a new finding, enables a number of methods for achieving encrypted computing. Certain researchers have even advocated for the use of cloud enclaves or trusted execution environments (TEE) to safeguard the AI process from start to finish. PyTorch and Tensorflow, two popular deep learning frameworks, support the creation of encrypted models. Encrypting health applications would bolster public confidence in AI’s adoption. As a consequence, including a range of sustainable healthcare modalities is strongly recommended in order to boost public trust in AI applications. As a result, the research community for federated learning has a special fondness for blockchain [70]. Due to the fact that existing deep learning application inference, training, deployment, and development procedures do not account for security issues, they are not accepted by society. After being trained on a great number of human faces “with” and “without” masks, a deep learning model is used to detect the probability of COVID-19 spreading. The use of technology to enhance people’s quality of life (QoL) is a critical component of sustainability. Modern healthcare facilities monitor a patient’s physiological, vital, and physical functions in real time or near real time, but the quantity and sophistication of monitoring technology accessible at home and in a medical institution are significantly different. As a consequence, providing patients with quality-of-life (QoL) support is hindered by the home environment. Patient’s quality-of-life monitoring at home becomes more essential as individuals approach the end of their lives since it enables people to maintain the traditional healthcare system [71]. It is important to have complete knowledge of the difficulties that older family members face in their later years. Even yet, little progress has been made

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

9

in the study of often-ignored side effects, that is, fatigue, nausea, condensed ability to maintain work, damaged relationships, depression, and an overall incapacity to accelerate productivity, meaningful daily lives. Treatments should aim to enhance patients’ overall well-being while also extending their lives. Patients experience physical, psychological, and social distress as a consequence of their illness and the aging process. These consequences are often ignored and undertreated in in-home elder care. Caregivers who estimate co-morbidity in older patients consistently miscalculate it, despite the fact that adverse clinical consequences are accurately assessed. This demonstrates how critical it is to get perspectives from both patients and caregivers. Patient-related outcome or patient-related experience data are seldom used in clinical intervention, making it challenging to collect QoL data from patients while they are still at home. AI has a bright future in the collection of QoL information from in-home patients because of recent developments in the medical AI industry. Artificial intelligence (AI) is capable of accurately analyzing data from video, audio, images, and healthcare IoT devices, thus assisting clinical decision-making. For instance, electronic health records, pathology, radiographic images, and genomes may all be automatically identified by various AI systems. However, as artificial intelligence (AI) advances and more decision-making is implemented through algorithmic “black boxes,” it is critical for an AI model to be explainable and trustworthy when used to analyze patient QoL data at home. Clinicians need semantic insight into the underlying ML procedures [72] if AI is to complement, if not completely replace, the intelligence of a medical practitioner. Making AI more “comprehensible” to healthcare practitioners may assist address this issue; for example, an oncologist may decide whether to trust QoL findings based on semantic evidence. Oncologists have acquired some qualitative, functional information, or what has been referred to as semantic interpretations, via the creation of Explainable AI (XAI), but utilizing XAI to collect QoL data from patients in their homes is a novel concept. By responding to queries such as “Would you trust the algorithm that suggests invasive surgery based on QoL data and tumor image classification?” XAI may assist oncologists in making more informed treatment choices. As shown in Fig. 4, this high-level example demonstrates the usage of XAI. When it originates to the future generation of AI, we think it will include multimodal capabilities to facilitate human-like ethics and prevent prejudice against datasets or inferencing. These characteristics will instill trust in stakeholders in the AI model while also ensuring the privacy and security of underlying datasets or models. Similarly, to how a physician interacts with another physician and shares their collective expertise when deciding on a treatment plan, the AI engine answers to human queries with relevant data and detailed explanations. This allows doctors to collaborate with their artificial intelligence counterparts. Figure 4 illustrates the interaction between a human doctor and an XAI as the doctor analyzes a melanoma skin cancer report produced by the XAI. This is the result reached after analyzing our skin cancer model in answer to a physician’s query. To convince the query engine of the results, the human physician may now interrogate it about different aspects of diagnostic processes and evidence. For instance, the XAI engine responds

10

S. Bharati et al.

Fig. 4 Basic scenario for bringing sustainability to IoHT-based XAI eco-systems

with the decision’s annotated region or bounding box. When a doctor treating a COVID-19 patient uploads a test X-ray image into a system with a XAI model, a diagnostic result is obtained. This is an example of XAI. When the test results are returned, the doctor will request further evidence. Four pieces of evidence were incorporated in the algorithm, according to XAI’s explainable model. The first step was to determine the most likely infection location. A second explanation for this is that the algorithm detected no similarities between this disease and any previously recognized pneumonia cases. Additionally, it demonstrated a 97.5% match to the COVID-19 symptom database. Finally, it was determined that the progress of the same patient’s days three, seven, and nine was essential for the transmission of the COVID-19 virus.

4 XAI in Healthcare for Different Purposes 4.1 Medical Perspectives The first question to ask from a medical standpoint is what makes AI-based clinical decision support different from other well-established diagnostic technologies, for example, improved laboratory tests. The results of both AI-based clinical decision support systems and improved lab tests can be used in CDASs, since the outputs of the schemes are verifiable and quantifiable. Other diagnostic techniques, such as

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

11

imaging, are similarly transparent to humans, and therefore are not called black-box approaches. However, using these methods, we are unable to explain the findings of any one test. As a consequence, it is apparent that there are two degrees of explainability in medicine. At the most fundamental level, explicitness enables us to understand how the system operates and how it arrives at its conclusions. We may use laboratory testing to show which inputs are needed for AIbased CDASs by evaluating the significance of features such as biochemical and biological processes. If a feature is explicable at the second level, we may deduce which qualities were essential for generating a certain forecast. For instance, if the distribution of characteristics outside the sample is unusual, the individual predictions may be analyzed for patterns that indicate a false prediction. Other diagnostic tests, on the other hand, will not have frequent access to the second level of explainability. Physicians’ explainability results will need to incorporate these developments (and patients). First-level explanations may be sufficient based on the therapeutic use case and the accompanying danger, while other use cases often need second-level explanations to safeguard patients. Currently, clinical validation is the most often stated need for a medical AI scheme. Oftentimes, explainability is missed until it is too late. To comply with regulatory standards and gain medical certification, medical AI schemes, particularly CDASs, whether or not they are driven by AI, must undergo a rigorous validation process [1]. After completing this validation process, it will be demonstrated that the CDAS schemes work in various real-world medical settings. It is essential to recognize the idea of clinical validation here. Prediction performance, which is frequently denoted as prediction accuracy, is a widely used performance measure. Numerous prediction accuracy measures exist, each tailored to a particular use case. These measures all, however, reflect a model’s capacity to generate accurate predictions and therefore its overall clinical utility. Increasing prediction accuracy and decreasing error rates are therefore critical models development goals. Artificial intelligence (AI) has been shown to significantly decrease error rates by up to 80% when compared to traditional methods. Regardless of their efforts, artificial intelligence (AI) systems will always have a margin of error due to the diversity of variables that may contribute to their inaccuracy. The first issue is that, due to the nature of medical datasets (e.g., noise or transcription errors), it is impossible to build an errorfree model. These errors are the consequence of a stochastic process. Thus, the possibility of making an incorrect positive or negative prediction is always present. Another major source of error is due to artificial intelligence’s prejudice. In an ideal scenario, the data used to train the AI tool would reflect the population on which it would be deployed. Achieving this optimal condition via thorough clinical validation and development using a variety of data sources is a critical goal of AI in the development of healthcare solutions. While this helps to mitigate AI bias, it is almost impossible to develop fully bias-free AI technologies. If bias exists, prediction errors will occur in patients who do not represent the training sample. Even if a completely validated, high-performing AI system is deployed, physicians and patients will continue to encounter random and systematic errors in the clinical setting. Thus, from a medical perspective, not just clinical validation, but also

12

S. Bharati et al.

explainability, is critical in the clinical setting. If an AI system and a human expert disagree, explainability enables the disagreement to be resolved regardless of who made the judgment error. It is critical to keep in mind that this works best when there is systematic error, such as when an artificial intelligence commits a mistake. When the tool and the physician agree or disagree, random errors are far more difficult to identify and more likely to go unnoticed. This scenario will be illustrated in detail under the ethical consequences. The results of explainability tests are often signified visually or with straightforward descriptions in everyday language. Both convey to doctors the factors that went into settling on a certain treatment strategy. In other words, physicians may utilize explainability to determine how well a system’s recommendations fit their clinical judgment and knowledge. This increases customer confidence in the system since they have more information to base their choice on. For instance, if the CDAS provides recommendations that are significantly different from what a physician anticipated, explainability may assist evaluate if the system’s parameters make clinical sense. By revealing the CDAS’s inner workings, explainability aids physicians in identifying false positives and negatives. If physicians find a flaw in the system, they may inform product designers so that they may work on resolving the issue. As a result, explainability may be a key element in determining whether AI-driven CDAS are adopted in clinical contexts where physicians lack full trust [73, 74]. It is critical to point out that any application of AI-based CDAS can affect a physician’s judgment. As a consequence, it will be critical to maintain detailed records of how recommendations were produced.

4.2 Patient Perspectives Patient-centered care concepts are in tension with artificial intelligence-based decision aids because of the question of explainability. The aim of patient-centered care is to be attentive to as well as respectful of each patient’s needs and values. Because it sees patients as full of life partners in the treatment procedure, it places a premium on patient control and choice over medical decisions. Collaborative decision-making is important for patient-centered care since it helps identify which therapies are suitable for particular patients [75, 76]. Patients and doctors participate in an open conversation during which the physician discusses the potential risks and benefits of different treatment choices and the patient communicates their priorities and values [77, 78]. Numerous evidence-based implements, including discussion aids [79], have been developed to help groups in making decisions. In contrast to patient choice aids, conversation aids are designed to be used [78, 80]. Patients may better understand their risks and results by incorporating well-established healthcare facts about their illnesses, and by synthesizing the accessible data, they can explore treatment choices and decide on a plan that fits their goals and priorities the most effectively. What if, rather of relying on well-established risk prediction algorithms, we used a data-driven approach that was both verifiable and unaccountable? Do you believe it will have an effect on the patient? In an effort to address these problems,

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

13

there has recently been an argument that “black-box medicine,” as it is often referred to, violates the basic ideals of patient-centered care. Doctors are unable to explain to patients how certain findings or recommendations were reached because they no longer fully grasp the decision aid’s internal workings and calculations. This issue may be solved by using explainability’s tailored conversation help for physicians and patients. By simulating the consequences on a patient’s health, the adoption of an explainable AI decision tool may help patients in becoming more conscientious of their treatment or lifestyle choices [81]. Explainability, it has been said, aids in providing a natural language or a visual representation explanation of the many factors that went into the concluding risk assessment. When evaluating system-derived probabilities and explanations, patients, on the other hand, depend on the clinician’s ability to convey and understand these explanations in an accurate and intelligible way. When applied properly, explainable artificial intelligence decision support schemes may help patients feel more informed and in charge of their health, as well as enhance their risk perceptions [81, 82]. As a consequence, patients’ willingness to engage in collaborative act and decisionmaking on risk-relevant data may be enhanced [82].

4.3 Technological Perspectives We will examine two issues from a technological perspective. To begin, let us discuss explainability methods and their use in medical AI research. When it comes to methodology, explainability may either be a built-in feature of an approximated or algorithm using other approaches [5, 83]. The concluding is essential for approaches, for example, ANN models, dubbed “black-box” models in the past. However, a variety of explanations for these predictions exist today [84]. Additionally, methods that simply approximate explainability [83] will often be less accurate than those that are intrinsically explainable. This is because many of today’s machine learning methods have very complex characteristics. It is essential to interpret the internal workings of artificial neurons, such as those found in ANNs, in order for humans to understand them. Intrinsically explainable methods offer a significant advantage over those that are not. However, these novel methods are often employed in combination with established techniques such as linear or logistic regression. In many application scenarios, traditional methods like neural networks (ANNs) perform worse than state-of-the-art technology [85]. As a consequence, inventors of CDASs are forced to choose between performance and explainability. As Rudin et al. [83] point out, in the minds of some, the problem is a misunderstanding of modeling methods rather than anything inherent in reality. Although Rudin et al. have highlighted the limits of approximation explainability approaches, given the difficulty of explaining machine learning models, specific approximating methods, rather than the notion of [83], may be beneficial. While we may argue qualitatively that intrinsic explainability is better to approximated explainability, no systematic quantitative comparisons have been conducted [86].

14

S. Bharati et al.

Nonetheless, newer algorithms like as ANNs are de facto favored for a wide variety of applications—and are often used in the development of AI products. Additionally, it is conceivable that newer methods perform better in certain instances. This necessitates a more critical assessment of explainability methods, both in terms of technical progress, such as method ranking and optimization for particular inputs, and in terms of explainability’s role from a multi-stakeholder viewpoint, as done in the current study. Explainability assists AI models in gaining user confidence in the models’ decisions. To illustrate, it is a good idea to exclude the possibility that meta-data is influencing prediction accuracy rather than actual data. To illustrate outside of medicine, when asked to categorize dogs into two groups—wolves and huskies—the predictor was determined only by whether or not the canines were in the snow [87]. Another term for this phenomenon is the “Clever Hans” phenomenon. In the medical sector, the phenomena of smart Hanses are also seen [88]. Mount Sinai Hospital researchers developed a model that was very effective in distinguishing high-risk people from those who were not. However, the tool’s performance dropped. Regrettably, the AI model never learned anything therapeutically valuable from the images. As in the prior example [20, 34], the prediction was generated using hardware meta-data connected with the particular X-ray equipment employed only at Mount Sinai to scan high-risk intensive care unit patients [89]. As a consequence, the system was able to determine just which imaging equipment was being used, not whether or whether people were in risk [20]. Because explainability techniques classify the Clever Hans predictors (snowy backdrop, hardware information) as prediction-relevant characteristics rather than domain-relevant data, explainability approaches enable inventers to detect these kinds of mistakes prior to clinical validation and certification [34]. This saves time and money on the project. Note that explainability techniques targeted at developers have different requirements than systems targeted at technologically unskilled end users like clinical physicians and patients. These approaches and representations may be more challenging for programmers to grasp.

4.4 The Legal Perspective Explainability in AI is a legal need; the question is to what extent this requirement is required legally. Transparency and traceability standards in healthcare are stricter than in other sectors, such as public administration [32, 90]. We have previously seen how methods such as supervised learning and deep learning may assist enhance healthcare. Diagnostic anomaly detection and pattern recognition are already changing clinical practice and healthcare standards. Permission and autonomy of patients must be addressed extensively in order to completely develop these prospects for saving lives and progressing patient outcomes via advances in disease detection and prevention. As a result, all applicable laws, regulations, and other legal requirements must be followed when it comes to data—its collection, storage, transfer, processing, and analysis. The law, as well as its interpretation

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

15

and implementation, must evolve in lockstep with technological advancements [91]. Even if all of these self-evident requirements are satisfied, the question of whether AI-driven solutions and tools need explanations persists. This implies that doctors and patients should be informed not just of the findings, but also of the characteristics and qualities upon which those findings were founded, as well as the important assumptions that underpin those findings. If new stakeholders are to be involved, are algorithms and models need to be understandable and explicable? From a Western legal viewpoint, we found three major categories of explainability: Three elements are required for medical device regulation: approval, certification, and informed consent as medical gadgets (as defined by the FDA and MDR), and responsibility. Personal medical data can be lawfully handled only with the individual’s consent. Since there are no uniform laws allowing for the use of personal information and data in AI applications, informed consent is the current industry standard [92]. Because authorization must be stated in advance, specifically, the purpose and objectives of the provided project must be specified, this presents an especially tough issue. Among the inherent benefits of AI is its ability to identify novel patterns and find novel biomarkers without needing preselection of features. This specific advantage may not be completely realized if utilized only for defined purposes, as informed consent requires. Individual and full information about these processes, as well as awareness of them, are needed by law in order to get informed consent for diagnostic treatments or interventions. AIbased decision aid necessitates describing the underlying processes and algorithms to each patient. As with consenting to an MR imaging therapy, the patient does not require to be informed of each detail, but he or she must be made aware of the basic ideas and, most crucially, the risks. Notably, the breadth of information for AI-based CDAS must be customized for each use case as well as will almost undoubtedly require clarification from legal authorities, since it is very difficult to define a priori. Beaudouin et al. [93] advocate for the establishment of a framework for assessing the “right” degree of explainability. Due to regulatory bodies, there have been delays in establishing certification and approval standards for explainable AI and its implications for product development and marketing. The FDA has released a discussion paper detailing how it supports the ongoing development and improvement of medical devices powered by artificial intelligence as part of its entire product lifecycle strategy (TPLC). Although explainability is not addressed, a sufficient degree of transparency of the product and technique directed at customers is needed [94]. This section is mostly concerned with the software’s features and their evolution through time. The MDR does not, however, stipulate that medical gadgets that include machine learning and artificial intelligence must be understandable to patients. Even yet, the need of accountability and transparency is firmly established here, and the growth of XAI may compel lawmakers and notified authorities to modify relevant laws and interpretations. Currently, the MDR and FDA demand explainability in a wide sense, that is, transparency, information for traceability, and explanation of machine learning/deep learning model development that is used to guide medical treatment. In the future, manufacturers of artificial intelligence-based medical equipment and software will

16

S. Bharati et al.

be expected to provide insight into the training and testing of models, as well as the data and overall development processes. It is also worth mentioning that the European Union’s General Data Protection Regulation (GDPR) is now debating whether explainable AI must be utilized in systems that handle patient data. Additionally, it is conceivable that the perplexing terminology employed today may be changed with something more understandable in the future. Another point of contention is whether or not patients should be informed that treatment recommendations generated by clinical decision support systems that include artificial intelligence may result in legal and litigation repercussions if the practitioner disregards them. According to Cohen, it is currently unclear in the United States how much machine learning/deep learning integration into clinical decision-making must be revealed to avoid liability [92]. Apart from the possibility of medical misuse lawsuit, doctors may be compelled to utilize a particular instrument that is comprehensible under contract and tort law [95]. The courts, however, will have the last say on this issue, and that judgment will likely come sooner rather than later as more AI-based systems are implemented. The legal implications of integrating artificial intelligence (AI) into healthcare are significant, requiring careful balancing to address the continuous conflict between development and regulation. Even with new cancer therapies or medications, AI-based decision aid has the potential to save lives, but standards and legal crash barriers are required to guarantee that patients’ rights and autonomy are not unintentionally infringed. Explainability is important in this context, and we think that performance is sufficient only if explainability cannot be given. In general, legal issues must be explained, and it has become necessary to open the black box, which will be a watershed moment for the use of artificial intelligence in medicine.

4.5 Ethical Implications As AI-powered technologies grow more ubiquitous in healthcare, it is more critical than ever to consider the ethical implications of this impending paradigm shift. Beauchamp and Childress’ “Principles of Biological Ethics” [96, 97] include four fundamental ideas that are often used and are well-suited for assessing biological ethical dilemmas: non-maleficence, autonomy, justice, and beneficence [96]. No other framework exists for bioethics, yet principles are widely accepted in both medical and scientific contexts because they are simple and effective. As a consequence, the next section assesses explainability using the four previously mentioned criteria. Explainability has implications for both patients’ and physicians’ autonomy [98]. Informed consent, which is a voluntary, typically written authorization given by a patient to a clinician to perform a particular medical act, is critical for preserving patients’ autonomy [99]. Complete disclosure of the nature and risks of medical treatments, as well as no interference with the patient’s free choice to undergo treatment, are the pillars of informed consent. The ethical consensus on whether or not disclosing the existence of a mystery algorithm in medical AI should be a

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

17

prerequisite of up-to-date consent is currently being developed. The doctor-patient and patients’ autonomy relationship may be jeopardized if an opaque AI system is not revealed, jeopardizing their trust and perhaps breaching therapeutic standards. The patient may contest the doctor’s advice if they later discover that it was based on an opaque AI system, and they may also want an explanation, which the clinician cannot provide if the system is opaque. Thus, the deployment of opaque medical AI may constitute a barrier to the provision of accurate information, jeopardizing informed consent. The autonomy-preserving function of informed consent must be safeguarded by acceptable ethical and explainability standards. In medical decisionmaking, consider how opaque artificial intelligence may encourage paternalism by restricting patients’ ability to communicate expectations and desires [99]. Sharing decision-making is only feasible when patients are presented with significant options, which can occur only when the patient has full autonomy. Patients will have fewer choices as opaque artificial intelligence (AI) takes over decision-making in medicine. The Opacity CDAS poses a unique issue in that it is unclear whether or not the model considers patients’ values and preferences. This scenario may be addressed via the use of “value-flexible” artificial intelligence, which provides patients with a range of options. We continue by stating that value-aware AI requires explainability. The AI system has a number of factors that the patient must be aware of in order to evaluate if the system’s objectives and weighting are consistent with their values. For example, the emphasis placed on “relief of pain” by patients may be at odds with AI systems programmed for “survival” [100]. Patients must have trust and autonomy in an AI system to follow its advice after a choice has been made [101]. The AI model must be transparent in order for this to be a viable choice. As a consequence, technologies that assist doctors in making important medical decisions must be morally acceptable from both the physician and patient viewpoints. While beneficence and non-maleficence are inextricably connected, they offer insight into distinct elements of their acts and motivations, such as their explainability. Physicians are driven by compassion to give their patients the finest treatment possible. Thus, doctors who use AI-based schemes are required to employ the tools in a method that maximizes the patient’s result. Medical practitioners must have access to all system capabilities in order to provide the best treatment possible to their patients. Physicians will be able to assess the system’s output if they understand more about it than how it operates in the clinical environment. Explainability through visuals or natural language explanations enables doctors to make confident clinical choices rather than relying only on a computerized output. They are capable of analytically evaluating the method’s outputs and drawing their own decisions regarding the findings’ reliability. As a consequence, they may alter their predictions and recommendations in light of new information. By using clinical judgment, physicians can not merely moderate the risk of arousing despair instilling or false hope, but they may also detect possibly ineffective therapies. Consider a situation in which a physician and an artificial intelligence system are at odds since this is unlikely to be resolved fast [101]. According to Grote et al. [101], there is inadequate epistemic justification for deference. They assert that when physicians are confronted with a black-box technology, clinical decision aid

18

S. Bharati et al.

may impede rather than enhance physician abilities. Doctors may be forced to perform “defensive medicine” in this situation, in which they blindly follow the machine’s suggestions in order to avoid being held or questioned accountable [101]. Physician autonomy would be gravely threatened in such a situation. Additionally, physicians will have a limited amount of time to investigate why their clinical judgment varies from that of the AI system. As a consequence, concentrating only on performance output is inadequate in the therapeutic context. Only healthcare professionals capable of making well-informed decisions about when and how to utilize an AI-powered CDAS can ensure the greatest potential result for all patients. Because beneficence cannot be fulfilled by any “black-box” application for the purpose of medical AI, it is impossible to envision how this might be accomplished. When the notion of non-maleficence is considered for the purpose of medical AI, the need of explainability becomes clearer. According to non-maleficence, physicians have a moral responsibility to avoid harming their patients, whether intentionally or via the use of excessive or inappropriate medical procedures. The physician may be ignorant of the causative processes underlying an AI-prescribed intervention; however, some have claimed that a morally acceptable medical AI black box uses only the most demonstrated maximum efficiency [102]. In medicine, it is still extremely common to rely on anecdotal or merely observed evidence to substantiate the efficacy of a specific treatment. This should not be used as an excuse to delay responding when sound clinical judgment requires it. According to the fairness principle, everyone should be able to benefit from medical advances without being subjected to unjustified discrimination on the basis of race, gender, or socioeconomic position [96]. However, certain AI systems defy this notion. Some researchers recently developed a medical AI scheme that discriminates against people based on their race [103]. Detecting and fixing these biases—a major source of unfairness in AI development and validation—may be helped in the initial phases of AI validation and development by explainability, for example, by identifying critical model features that signal a bias. It is critical to educate the relevant stakeholder groups about the bias risk and the potential consequences for people’s well-being before explainability can perform its role. It is alluring to prioritize accuracy above investing time and money in developing AI that makes sense. To maximize the promise of AI-powered decision-assisted systems, however, both doctors and developers must be cognizant of the new tools’ potential flaws and limitations. As a result, explainability becomes a legal requirement for the advancement and use of AI-based clinical decision-assisted systems.

5 Conclusion In this chapter, we describe the role of explainable artificial intelligence in clinical decision-assisted systems from the perspectives of technology, law, medicine, and patients. This chapter also depicts the role of IoHT in XAI framework. Consequently, we have shown that explainability is a complex notion with far-

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

19

reaching consequences for the many stakeholder groups involved. AI-based medical system offers a new set of challenges for developers, medical practitioners, and policymakers, since it needs a rethinking of responsibilities and roles. According to this chapter results, explainability is a necessary condition for resolving these problems sustainably and in accordance with professional norms and values. It is critical to keep in mind that shifting CDAS toward opaque algorithms may inadvertently result in a return to paternalistic care concepts that reduce patients to passive spectators. Medical practitioners may become enslaved to the tool’s output in order to escape medical and legal repercussions, creating a new kind of medicine. Finally, opaque systems may result in resource misallocation, violating the concept of fair distribution. In this chapter, we claimed that people should be at the center of treatment and should be allowed to make informed and autonomous health decisions with the help of physicians. Explanability may also contribute to more fair resource distribution. We determine that excluding explainability from CDASs violates basic medical ethical norms, which may have severe consequences for the future patient and public well-being. More education is required to inform healthcare practitioners, developers, and policymakers about the limitations and challenges of medical AI’s opaque algorithms, as well as to foster cross-disciplinary cooperation to address these concerns.

References 1. Higgins, D., & Madai, V. I. (2020). From bit to bedside: A practical framework for artificial intelligence product development in healthcare. Advanced Intelligent Systems, 2(10), 2000052. 2. Robel, M. R. A., Bharati, S., Podder, P., & Mondal, M. R. H. (2020). IoT driven healthcare monitoring system. Fog, Edge, and Pervasive Computing in Intelligent IoT Driven Applications, 161–176. 3. Singh, R. P., Javaid, M., Haleem, A., Vaishya, R., & Ali, S. (2020, July 1). Internet of Medical Things (IoMT) for orthopaedic in COVID-19 pandemic: Roles, challenges, and applications. Journal of Clinical Orthopaedics and Trauma, 11(4), 713–717. 4. Bharati, S., & Hossain Mondal, M. R. (2021). 12 Applications and challenges of AI-driven IoHT for combating pandemics: a review. In A. Khamparia, R. Hossain Mondal, P. Podder, B. Bhushan, V. H. C. D. Albuquerque, & S. Kumar (Eds.), Computational intelligence for managing pandemics (pp. 213–230). De Gruyter. https://doi.org/10.1515/9783110712254012 5. Yang, G., Ye, Q., & Xia, J. (2022). Unbox the black-box for the medical explainable AI via multi-modal and multi-centre data fusion: A mini-review, two showcases and beyond. Information Fusion, 77, 29–52. https://doi.org/10.1016/j.inffus.2021.07.016 6. Bharati, S., Mondal, M. R. H., & Podder, P. (2023). A review on explainable artificial intelligence for healthcare: Why, how, and when? IEEE Transactions on Artificial Intelligence. https://doi.org/10.1109/TAI.2023.3266418 7. Jiang, F., Jiang, Y., Zhi, H., Dong, Y., Li, H., Ma, S., Wang, Y., Dong, Q., Shen, H., & Wang, Y. (2017). Artificial intelligence in healthcare: Past, present and future. Stroke and Vascular Neurology, 2(4), 230. 8. Panch, T., Mattie, H., & Celi, L. A. (2019). The “inconvenient truth” about AI in healthcare. NPJ Digital Medicine, 2(1), 1–3.

20

S. Bharati et al.

9. Yu, K.-H., Beam, A. L., & Kohane, I. S. (2018). Artificial intelligence in healthcare. Nature Biomedical Engineering, 2(10), 719–731. 10. Shen, D., Wu, G., & Suk, H.-I. (2017). Deep learning in medical image analysis. Annual Review of Biomedical Engineering, 19, 221–248. 11. Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F., Ghafoorian, M., Van Der Laak, J. A., Van Ginneken, B., & Sánchez, C. I. (2017). A survey on deep learning in medical image analysis. Medical Image Analysis, 42, 60–88. 12. Ker, J., Wang, L., Rao, J., & Lim, T. (2017). Deep learning applications in medical image analysis. IEEE Access, 6, 9375–9389. 13. Dilsizian, S. E., & Siegel, E. L. (2014). Artificial intelligence in medicine and cardiac imaging: Harnessing big data and advanced computing to provide personalized medical diagnosis and treatment. Current Cardiology Reports, 16(1), 441. 14. Patel, V. L., Shortliffe, E. H., Stefanelli, M., Szolovits, P., Berthold, M. R., Bellazzi, R., & Abu-Hanna, A. (2009). The coming of age of artificial intelligence in medicine. Artificial Intelligence in Medicine, 46(1), 5–17. 15. Jha, S., & Topol, E. J. (2016). Adapting to artificial intelligence: Radiologists and pathologists as information specialists. JAMA, 316(22), 2353–2354. 16. Strickland, E. (2019). IBM Watson, heal thyself: How IBM overpromised and underdelivered on AI health care. IEEE Spectrum, 56(4), 24–31. 17. Weingart, N. S., Wilson, R. M., Gibberd, R. W., & Harrison, B. (2000). Epidemiology of medical error. BMJ, 320(7237), 774–777. 18. Graber, M. L., Franklin, N., & Gordon, R. (2005). Diagnostic error in internal medicine. Archives of Internal Medicine, 165(13), 1493–1499. 19. Winters, B., Custer, J., Galvagno, S. M., Colantuoni, E., Kapoor, S. G., Lee, H., Goode, V., Robinson, K., Nakhasi, A., & Pronovost, P. (2012). Diagnostic errors in the intensive care unit: A systematic review of autopsy studies. BMJ Quality & Safety, 21(11), 894–902. 20. Podder, P., Khamparia, A., Mondal, M. R. H., Rahman, M. A., & Bharati, S. (2021). Forecasting the spread of COVID-19 and ICU requirements. International Journal of Online and Biomedical Engineering (iJOE), 17(5), 81. 21. Neill, D. B. (2013). Using artificial intelligence to improve hospital inpatient care. IEEE Intelligent Systems, 28(2), 92–95. 22. Clement, J. C., Ponnusamy, V., Sriharipriya, K. C., & Nandakumar, R. (2021, March 26). A survey on mathematical, machine learning and deep learning models for COVID-19 transmission and diagnosis. IEEE Reviews in Biomedical Engineering, 15, 325–340. 23. Bharati, S., Robel, M. R. A., Rahman, M. A., Podder, P., & Gandhi, N. (2019). Comparative performance exploration and prediction of fibrosis, malign lymph, metastases, normal lymphogram using machine learning method (pp. 66–77). Springer. 24. Miller, R. A. (1994). Medical diagnostic decision support systems—Past, present, and future: A threaded bibliography and brief commentary. Journal of the American Medical Informatics Association, 1(1), 8–27. 25. Berner, E. S. (2007). Clinical decision support systems (Vol. 233). Springer. 26. Schultheiss, M., Schmette, P., Bodden, J., Aichele, J., Müller-Leisse, C., Gassert, F. G., Gassert, F. T., Gawlitza, J. F., Hofmann, F. C., & Sasse, D. (2021). Lung nodule detection in chest X-rays using synthetic ground-truth data comparing CNN-based diagnosis to human performance. Scientific Reports, 11(1), 1–10. 27. Podder, P., Bharati, S., Rahman, M. A., & Kose, U. (2021). Transfer learning for classification of brain tumor. In Deep learning for biomedical applications (pp. 315–328). CRC Press. 28. Elaziz, M. A., Hosny, K. M., Salah, A., Darwish, M. M., Lu, S., & Sahlol, A. T. (2020, June 26). New machine learning method for image-based diagnosis of COVID-19. PLoS One, 15(6), e0235187. 29. Munnangi, A. K., Sekaran, R., Rajeyyagari, S., Ramachandran, M., Kannan, S., & Bharati, S. (2022, June). Nonlinear cosine neighborhood time series-based deep learning for the prediction and analysis of COVID-19 in India. Wireless Communications and Mobile Computing, 29, 2022.

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

21

30. Jain, D. K., Lakshmi, K. M., Varma, K. P., Ramachandran, M., & Bharati, S. (2022, May 27). Lung cancer detection based on Kernel PCA-convolution neural network feature extraction and classification by fast deep belief neural network in disease management using multimedia data sources. Computational Intelligence and Neuroscience, 2022, 1–12. 31. Bharati, S., Podder, P., & Mondal, M. R. H. (2020). Hybrid deep learning for detecting lung diseases from X-ray images. Informatics in Medicine Unlocked, 20, 100391. 32. Shen, L., Margolies, L. R., Rothstein, J. H., Fluder, E., McBride, R., & Sieh, W. (2019, August 29). Deep learning to improve breast cancer detection on screening mammography. Scientific Reports, 9(1), 1–2. 33. Bharati, S., Podder, P., & Mondal, M. R. H. (2020). Diagnosis of polycystic ovary syndrome using machine learning algorithms (pp. 1486–1489). IEEE. 34. Pathan, S., Siddalingaswamy, P. C., & Ali, T. (2021, June 1). Automated detection of COVID19 from chest X-ray scans using an optimized CNN architecture. Applied Soft Computing, 104, 107238. 35. Bharati, S., Podder, P., Mondal, M., & Prasath, V. B. (2021). Medical imaging with deep learning for COVID-19 diagnosis: A comprehensive review. International Journal of Computer Information Systems and Industrial Management Applications, 13, 91–112. 36. Feng, Y., Wang, X., & Zhang, J. (2021, April 13). A heterogeneous ensemble learning method for neuroblastoma survival prediction. IEEE Journal of Biomedical and Health Informatics., 26(4), 1472–1483. 37. Szolovits, P., & Pauker, S. G. (1994). Categorical and probabilistic reasoning in medicine revisited. Artificial Intelligence, 59, 167–180. 38. Han, S., Kelly, E., Nikou, S., & Svee, E.-O. (2021). Aligning artificial intelligence with human values: Reflections from a phenomenological perspective. AI & Society, 1–13. 39. Rajkomar, A., Dean, J., & Kohane, I. (2019). Machine learning in medicine. New England Journal of Medicine, 380(14), 1347–1358. 40. Khamparia, A., Bharati, S., Podder, P., Gupta, D., Khanna, A., Phung, T. K., & Thanh, D. N. H. (2020). Diagnosis of breast cancer based on modern mammography using hybrid transfer learning. Multidimensional Systems and Signal Processing, 32, 747–765. 41. Polat, Ç., Karaman, O., Karaman, C., Korkmaz, G., Balcı, M. C., & Kelek, S. E. (2021, January 1). COVID-19 diagnosis from chest X-ray images using transfer learning: Enhanced performance by debiasing dataloader. Journal of X-ray Science and Technology, 29(1), 19– 36. 42. Bharati, S., Rahman, M. A., & Podder, P. (2018). Breast cancer prediction applying different classification algorithm with comparative analysis using WEKA. In 2018 4th international conference on electrical engineering and information & communication technology (iCEEiCT), Dhaka, Bangladesh (pp. 581–584). IEEE. https://doi.org/10.1109/ CEEICT.2018.8628084 43. van der Velden, B. H., Kuijf, H. J., Gilhuijs, K. G., & Viergever, M. A. (2022, May 4). Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Medical Image Analysis, 79, 102470. 44. Kobyli´nska, K., Orłowski, T., Adamek, M., & Biecek, P. (2022, February 12). Explainable machine learning for lung cancer screening models. Applied Sciences, 12(4), 1926. 45. Silva, F., Pereira, T., Neves, I., Morgado, J., Freitas, C., Malafaia, M., Sousa, J., Fonseca, J., Negrão, E., Flor de Lima, B., & Correia da Silva, M. (2022, March 16). Towards machine learning-aided lung cancer clinical routines: Approaches and open challenges. Journal of Personalized Medicine, 12(3), 480. 46. Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press. 47. Tomczak, K., Czerwi´nska, P., & Wiznerowicz, M. (2015). The Cancer Genome Atlas (TCGA): An immeasurable source of knowledge. Contemporary Oncology, 19(1A), A68. 48. Sudlow, C., Gallacher, J., Allen, N., Beral, V., Burton, P., Danesh, J., Downey, P., Elliott, P., Green, J., & Landray, M. (2015). UK biobank: An open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Medicine, 12(3), e1001779.

22

S. Bharati et al.

49. Obermeyer, Z., & Emanuel, E. J. (2016). Predicting the future—Big data, machine learning, and clinical medicine. The New England Journal of Medicine, 375(13), 1216. 50. Krause, J., Gulshan, V., Rahimy, E., Karth, P., Widner, K., Corrado, G. S., Peng, L., & Webster, D. R. (2018). Grader variability and the importance of reference standards for evaluating machine learning models for diabetic retinopathy. Ophthalmology, 125(8), 1264– 1272. 51. Kirby, J. C., Speltz, P., Rasmussen, L. V., Basford, M., Gottesman, O., Peissig, P. L., Pacheco, J. A., Tromp, G., Pathak, J., & Carrell, D. S. (2016). PheKB: A catalog and workflow for creating electronic phenotype algorithms for transportability. Journal of the American Medical Informatics Association, 23(6), 1046–1052. 52. Holzinger, A., Malle, B., Saranti, A., & Pfeifer, B. (2021). Towards multi-modal causability with Graph Neural Networks enabling information fusion for explainable AI. Information Fusion, 71, 28–37. 53. Hossain, M. S., Muhammad, G., & Alamri, A. (2019). Smart healthcare monitoring: A voice pathology detection paradigm for smart cities. Multimedia Systems, 25(5), 565–575. 54. Bharati, S., Podder, P., Mondal, M. R. H., & Paul, P. K. (2021). Applications and challenges of cloud integrated IoMT. In Cognitive internet of medical things for smart healthcare (pp. 67–85). Springer. 55. Lim, G., Bellemo, V., Xie, Y., Lee, X. Q., Yip, M. Y. T., & Ting, D. S. W. (2020). Different fundus imaging modalities and technical factors in AI screening for diabetic retinopathy: A review. Eye and Vision, 7(1), 1–13. 56. Rahman, M. A., Hossain, M. S., Showail, A. J., Alrajeh, N. A., & Alhamid, M. F. (2021). A secure, private, and explainable IoHT framework to support sustainable health monitoring in a smart city. Sustainable Cities and Society, 103083. 57. Hossain, M. S., & Muhammad, G. (2016). Cloud-assisted industrial internet of things (iiot)– enabled framework for health monitoring. Computer Networks, 101, 192–202. 58. Shorfuzzaman, M., Hossain, M. S., & Alhamid, M. F. (2021). Towards the sustainable development of smart cities through mass video surveillance: A response to the COVID-19 pandemic. Sustainable Cities and Society, 64, 102582. 59. Hossain, M. S., Muhammad, G., Abdul, W., Song, B., & Gupta, B. B. (2018). Cloudassisted secure video transmission and sharing framework for smart cities. Future Generation Computer Systems, 83, 596–606. 60. Alamri, A., Hassan, M. M., Hossain, M. A., Al-Qurishi, M., Aldukhayyil, Y., & Hossain, M. S. (2014). Evaluating the impact of a cloud-based serious game on obese people. Computers in Human Behavior, 30, 468–475. 61. Chen, D., Wawrzynski, P., & Lv, Z. (2020). Cyber security in smart cities: A review of deep learning-based applications and case studies. Sustainable Cities and Society, 102655. 62. Hossain, M. S., & Muhammad, G. (2020). Deep learning based pathology detection for smart connected healthcares. IEEE Network, 34(6), 120–125. 63. Amin, S. U., Hossain, M. S., Muhammad, G., Alhussein, M., & Rahman, M. A. (2019). Cognitive smart healthcare for pathology detection and monitoring. IEEE Access, 7, 10745– 10753. 64. Muhammad, G., Hossain, M. S., & Kumar, N. (2020). EEG-based pathology detection for home health monitoring. IEEE Journal on Selected Areas in Communications, 39(2), 603– 610. 65. Bharati, S., Podder, P., & Raihan-Al-Masud, M. (2018). EEG eye state prediction and classification in order to investigate human cognitive state. 66. Chen, M., Hao, Y., Hu, L., Hossain, M. S., & Ghoneim, A. (2018). Edge-CoCaCo: Toward joint optimization of computation, caching, and communication on edge cloud. IEEE Wireless Communications, 25(3), 21–27. 67. Robel, M. R. A., Bharati, S., Podder, P., Raihan-Al-Masud, M., & Mandal, S. (2019). Fault tolerance in cloud computing-an algorithmic approach. In International conference on innovations in bio-inspired computing and applications (pp. 307–316). Springer.

Explainable Artificial Intelligence (XAI) with IoHT for Smart Healthcare: A Review

23

68. Kim, H., Park, J., Bennis, M., & Kim, S.-L. (2019). Blockchained on-device federated learning. IEEE Communications Letters, 24(6), 1279–1283. 69. Cha, J., Singh, S. K., Kim, T. W., & Park, J. H. (2021). Blockchain-empowered cloud architecture based on secret sharing for smart city. Journal of Information Security and Applications, 57, 102686. 70. ur Rehman, M. H., Salah, K., Damiani, E., & Svetinovic, D. (2020). Towards blockchainbased reputation-aware federated learning. 71. Chen, M., Yang, J., Hu, L., Hossain, M. S., & Muhammad, G. (2018). Urban healthcare big data system based on crowdsourced and cloud-based air quality indicators. IEEE Communications Magazine, 56(11), 14–20. 72. Qian, S., Zhang, T., Xu, C., & Hossain, M. S. (2015). Social event classification via boosted multimodal supervised latent dirichlet allocation. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 11(2), 1–22. 73. Cutillo, C. M., Sharma, K. R., Foschini, L., Kundu, S., Mackintosh, M., & Mandl, K. D. (2020). Machine intelligence in healthcare—Perspectives on trustworthiness, explainability, usability, and transparency. NPJ Digital Medicine, 3(1), 1–5. 74. Tonekaboni, S., Joshi, S., McCradden, M. D., & Goldenberg, A. (2019). What clinicians want: Contextualizing explainable machine learning for clinical end use. 75. Barry, M. J., & Edgman-Levitan, S. (2012). Shared decision making—The pinnacle patientcentered care. 76. Kunneman, M., Montori, V. M., Castaneda-Guarderas, A., & Hess, E. P. (2016). What is shared decision making? (and what it is not). Academic Emergency Medicine, 23(12), 1320– 1324. 77. O’Neill, E. S., Grande, S. W., Sherman, A., Elwyn, G., & Coylewright, M. (2017). Availability of patient decision aids for stroke prevention in atrial fibrillation: A systematic review. American Heart Journal, 191, 1–11. 78. Noseworthy, P. A., Brito, J. P., Kunneman, M., Hargraves, I. G., Zeballos-Palacios, C., Montori, V. M., & Ting, H. H. (2019). Shared decision-making in atrial fibrillation: Navigating complex issues in partnership with the patient. Journal of Interventional Cardiac Electrophysiology, 56(2), 159–163. 79. Dobler, C. C., Sanchez, M., Gionfriddo, M. R., Alvarez-Villalobos, N. A., Ospina, N. S., Spencer-Bonilla, G., Thorsteinsdottir, B., Benkhadra, R., Erwin, P. J., & West, C. P. (2019). Impact of decision aids used during clinical encounters on clinician outcomes and consultation length: A systematic review. BMJ Quality & Safety, 28(6), 499–510. 80. Noseworthy, P. A., Kaufman, E. S., Chen, L. Y., Chung, M. K., Elkind, M. S. V., Joglar, J. A., Leal, M. A., McCabe, P. J., Pokorney, S. D., & Yao, X. (2019). Subclinical and devicedetected atrial fibrillation: Pondering the knowledge gap: A scientific statement from the American Heart Association. Circulation, 140(25), e944–e963. 81. Politi, M. C., Dizon, D. S., Frosch, D. L., Kuzemchak, M. D., & Stiggelbout, A. M. (2013). Importance of clarifying patients’ desired role in shared decision making to match their level of engagement with their preferences. BMJ, 347, f7066. 82. Stacey, D., Légaré, F., Lewis, K., Barry, M. J., Bennett, C. L., Eden, K. B., Holmes-Rovner, M., Llewellyn-Thomas, H., Lyddiatt, A., & Thomson, R. (2017). Decision aids for people facing health treatment or screening decisions. Cochrane Database of Systematic Reviews, (4). 83. Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206–215. 84. Samek, W., Montavon, G., Vedaldi, A., Hansen, L. K., & Müller, K.-R. (2019). Explainable AI: Interpreting, explaining and visualizing deep learning (Vol. 11700). Springer Nature. 85. Mondal, M. R. H., Bharati, S., & Podder, P. (2021). CO-IRv2: Optimized InceptionResNetV2 for COVID-19 detection from chest CT images. PLoS One, 16(10), e0259179. https://doi.org/ 10.1371/journal.pone.0259179 86. Islam, S. R., Eberle, W., & Ghafoor, S. K. (2020). Towards quantification of explainability in explainable artificial intelligence methods.

24

S. Bharati et al.

87. Samek, W., Montavon, G., Lapuschkin, S., Anders, C. J., & Müller, K.-R. (2020). Toward interpretable machine learning: Transparent deep neural networks and beyond. arXiv eprints:arXiv-2003. 88. Lapuschkin, S., Wäldchen, S., Binder, A., Montavon, G., Samek, W., & Müller, K.-R. (2019). Unmasking Clever Hans predictors and assessing what machines really learn. Nature Communications, 10(1), 1–8. 89. Zech, J. R., Badgeley, M. A., Liu, M., Costa, A. B., Titano, J. J., & Oermann, E. K. (2018). Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLoS Medicine, 15(11), e1002683. 90. Olsen, H. P., Slosser, J. L., Hildebrandt, T. T., & Wiesener, C. (2019). What’s in the box? The legal requirement of explainability in computationally aided decision-making in public administration. 91. Schönberger, D. (2019). Artificial intelligence in healthcare: A critical analysis of the legal and ethical implications. International Journal of Law and Information Technology, 27(2), 171–203. 92. Cohen, I. G. (2019). Informed consent and medical artificial intelligence: What to tell the patient? The Georgetown Law Journal, 108, 1425. 93. Maxwell, W., Beaudouin, V., Bloch, I., Bounie, D., Clémençon, S., d’Alché-Buc, F., Eagan, J., Mozharovskyi, P., & Parekh, J. (2020). Identifying the ‘right’ level of explanation in a given situation. 94. Food and Drug Administration. (2019). Proposed regulatory framework for modifications to artificial intelligence/machine learning (AI/ML)-based software as a medical device (SaMD). 95. Hacker, P., Krestel, R., Grundmann, S., & Naumann, F. (2020). Explainable AI under contract and tort law: Legal incentives and technical challenges. Artificial Intelligence and Law, 1–25. 96. Beauchamp, T. L., & Childress, J. F. (2001). Principles of biomedical ethics. Oxford University Press. 97. Gillon, R. (2015). Defending the four principles approach as a good basis for good medical practice and therefore for good medical ethics. Journal of Medical Ethics, 41(1), 111–116. 98. Spencer-Bonilla, G., Thota, A., Organick, P., Ponce, O. J., Kunneman, M., Giblon, R., Branda, M. E., Sivly, A. L., Behnken, E., & May, C. R. (2020). Normalization of a conversation tool to promote shared decision making about anticoagulation in patients with atrial fibrillation within a practical randomized trial of its effectiveness: A cross-sectional study. Trials, 21, 1–10. 99. Faden, R. R., & Beauchamp, T. L. (1986). A history and theory of informed consent. Oxford University Press. 100. McDougall, R. J. (2019). Computer knows best? The need for value-flexibility in medical AI. Journal of Medical Ethics, 45(3), 156–160. 101. Grote, T., & Berens, P. (2020). On the ethics of algorithmic decision-making in healthcare. Journal of Medical Ethics, 46(3), 205–211. 102. London, A. J. (2019). Artificial intelligence and black-box medical decisions: Accuracy versus explainability. Hastings Center Report, 49(1), 15–21. 103. Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453.

ARIMA and Predicted Geospatial Distribution of COVID-19 in India Prisilla Jayanthi and Iyyanki MuraliKrishna

1 Introduction ARIMA (AR) is a time series method and well-known approach to machine learning. Machine learning deals with the various predictions and classification techniques. AR model is found to be a statistical approach in machine learning for the prediction of various issues. Maleki et al. (2020) used time series models to predict and forecast the future confirmed cases (CC) and found that the symmetry of error distribution is very essential [1]. The proposed time series models were well suited for ordinary Gaussian and symmetry models for COVID-19 (CV-19) datasets. Duan and Zhang (2020) proposed an AR model to analyze two datasets of CV-19 in Japan and South Korea (January 20–April 26, 2020) and predicted cases for the 7 days (April 27–May 3, 2020) [2]. Khan and Gupta (2020) developed an AR model and a nonlinear autoregressive (NAR) NN for comparing the accuracy of predicted models. The models showed an increasing pattern of 1500 cases per day of the CV-19 (April 2020) dataset [3]. Ceylan (2020) developed the following models: Italy, AR (0,2,1); Spain, AR (1,2,0); France, AR (0,2,1), and few in Europe. The models predicted the epidemiological trend for the CV-19 prevalence dataset (February 21 to April 15, 2020) [4]. Sahai et al. (2020) used the time series to analyze the CV-19 data by forecasting the highly affected top five countries. The analysis proves that India and Brazil will reach 1.38 million and 2.47 million, respectively, while the USA will reach 4.29 million by July 31, 2020 [5]. Benvenuto et al. (2020) proposed an econometric P. Jayanthi () Faculty Associate, Department of CSE, Ecole Centrale School of Engineering, Mahindra University, Jeedimetla, Hyderabad, India I. MuraliKrishna Former Dr. Raja Ramanna Distinguished Fellow DRDO and Director R&D JNT University, Smart Village Movement In Alliance With Berkeley Haas, Hyderabad, India © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_2

25

26

P. Jayanthi and I. MuraliKrishna

model named the AR model that would predict the prevalence and incidence trend of CV-19. Thereafter, it was stated that any public health and infection control system could use the model and can daily construct a reliable forecast for the CV-19 epidemic [6]. Feroze (2020) proposed the Bayesian structural time series (BSTS) and AR models to explore the temporal dynamics of CV-19 data in top five affected countries from March to June 2020 [7]. As the cases were increasing due to a lack of sanitation awareness, Singh (2020) pointed out, that in most rural areas of India, more number of households are extremely exposed to CV-19 and sensitive to insufficient and poor availability of clean water, sanitation, and healthcare system in one’s surroundings [8]. The authors (Lai and Dzombak 2020) desired to have an efficient and adaptable statistical predicting method with high interpretability for their study. The ARIMA model provides more accurate forecasts and also reliability than other statistical techniques. Later the authors suggested the use of the ARIMA-based forecasting model as an interpretable model, reliable near-term, location-specific temperature and precipitation forecasts for concern of climate change in the applications of civil and environmental engineering [9].

2 Discussion and Results This study is an extension of the previous work where ARIMA (1,1,0) was modeled for predicting. The previous work (Iyyanki and Prisilla 2020) included a comparative study of two predictive models of time series, AR and KALMAN filter. Both the models were used on the stationary time series datasets. But, it was interesting to observe that the AR model gave better results than the KALMAN filter model for the CV-19 dataset [10]. In this study, the AR models were implemented and analyzed for prediction in four states, namely AP, Delhi, Maharashtra, and Tamil Nadu along with the confirmed and death cases of the entire country (India). Before the study begins, let us learn a little about what an interpretable model is all about. Interpretable models can interpret; certain models that come under this category are linear regression (LR), logistic regression, KNN, k-means clustering, and decision trees. LR models work better when the predictors are not correlated and are independent of each other. “Auto Regressive” word in ARIMA is an LR model that uses its lags as predictors. Hence, the study involves an Interpretable model, i.e., ARIMA model that is obtained as a result [11–13]. The cases (Table 1) were considered from March 12, 2020, to August 15, 2020, and predicted for September 15, October 15, and November 15, 2020 (Table 2) using ARIMA (4,0,2) and ARIMA (2,0,2) models for the confirmed and death cases in India, respectively. The graphs (Fig. 1) represent the total number of confirmed and death cases that are obtained by the time series approach. The daily increase in both the cases was recorded from March 12, 2020, to August 15, 2020, and examined carefully and it was found that the highest increase

India-Con 110 11,933 81,970 3,32,424 9,36,181 25,87,634

India-DT 2 392 2,649 9,520 24,309 50,454

Source: MOHFW site con confirm cases, DT death cases

Dates/states Mar 15 Apr 15 May 15 June 15 July 15 Aug 15

Table 1 Monthly wise confirmed cases MH-Con 32 2,337 27,524 1,07,958 2,67,665 5,84,754

MH-DT 0 160 1,019 3,950 10,695 19,749

TN-Con 1 1,173 9,674 44,661 1,47,324 3,32,105

TN-DT 0 11 66 435 2,099 5,641

AP-Con 1 473 2,205 6,163 33,019 2,81,817

AP-DT 0 9 48 84 408 2,562

DE-Con 7 1,512 8,470 41,182 1,15,346 1,51,928

DE-DT 1 28 115 1,327 3,446 4,188

ARIMA and Predicted Geospatial Distribution of COVID-19 in India 27

28

P. Jayanthi and I. MuraliKrishna

Table 2 The prediction table for confirmed and death cases in India for the month of September, October, and November, 2020 Predicted ConfirmARIMA(4,0,2)

Date

India Confirmed Cases

15-Mar-20

110

India Death Cases 2

15-Apr-20

12,759

392

15-May-20

81,970

2,649

15-Jun-20

3,32,424

9,520

15-Jul-20

9,36,181

24,309

15-Aug-20

25,26,192

50,454

Predicted Death ARIMA(2,0,2)

15-Sept-20

43,19,625

78,233

15-Oct-20

60,06,375

1,06,461

15-Nov-20

75,61,324

1,33,569

Fig. 1 The graph of India’s confirm and death cases (Table 1)

in CCs (66,999 cases) was on August 13, 2020, and that of death cases (2003 cases) was on June 17, 2020, as shown in the graph (Fig. 2). There was communal spread in some states which lead to a rise in cases in India with an abrupt increase.

2.1 Graphs for Confirmed and Death Cases, India The graphs (Fig. 1) indicate the increase of the COVID-19 CC and death cases from July 15 to August 15, 2020. There was a slow rise of cases during the lockdown period, but as the lockdown was relaxed, and then the cases were spotted high.

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

29

Fig. 2 Day-wise increase: (a) confirmed and (b) death cases in India

2.2 Graphs for Daily Increase in Confirm and Death Cases, India ARIMA (p,d,q) model is trained with the given dataset, where p represents autocorrelation, order of difference (d), and the moving average (q) giving different results based on the combinations. The trend of the prediction curve for India’s confirmed cases shows an increase up to 60 lakhs by October 2020 with ARIMA (4,0,2); ARIMA (2,0,3), and with ARIMA (4,0,0) it gave 59 lakhs.

2.3 Prediction of Confirmed Cases in India The prediction of CC for India with ARIMA (4,0,2) is shown in the graph (Fig. 3) with red dashed lines. Figure 4a, b show the vertical axis of the graph denotes the AC/PAC of CC with a yielding range from −1.00 to +1.00 with a rise of 0.50, and the horizontal axis denotes the lag ranging from 0 to 40 with a rise of 10. The estimated coordinate blue dots are represented with the calculated values of CC as ACF and PACF. Similarly, Fig. 4c–f indicate ACF and PACF for cases of daily increase in confirmed and predicted CC, respectively.

2.4 Prediction of Death Cases in India The death cases are trained for ARIMA (1,0,1) and ARIMA (2,0,2) models which produced the results with fifty one thousand and one lakh, respectively. The accuracy of a model was tested by comparing the actual values with the predicted values. Here, three performance standards, namely, RMSE, MAE, and MAPE were applied

30

P. Jayanthi and I. MuraliKrishna

Fig. 3 The prediction graph of CC in India

for the predicting accuracy of AR models [4]. The standards are expressed using the notation as follows:   n   RMSE = 1/n (ex )2

(1)

x=1

MAE = 1/n

n 

| ex |

(2)

   ex    x=1  yx 

(3)

x=1

and n MAPE = 100%/n

where ex represents the difference between the observed and estimated values of cases. yx , represents the observed values, and n represents the number of observations made. By comparing the calculated values of RMSE, MAE, and MAPE, the least value of the model is presumed to be the best model for prediction. In Fig. 5, the prediction for death cases in India was obtained by ARIMA (2,0,2) model shown with the dash lines. The trend appears to increase to 1 lakh by October and 1.3 lakhs by the middle of November, 2020. It was observed that India’s

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

31

Fig. 4 ACF and PACF graph of (a, b) CC, (c, d) CC day-wise increase (e, f) predicted CC of overall India

death rate has a minimum of 2.38% because of native immunity in the food style consumed; though India stands third in position in the world. The graph for the death cases in India, daily increase in cases, and predicted cases are shown in Fig. 6 with the vertical axis yielding the range from negative 1.00 to 1.00 with increments of 0.50. Table 3 is obtained with the computation of Eqs. (1)–(3), and ARIMA (4,0,2) and ARIMA (2,0,2) are found to best fit models for confirm and death cases, respectively. The model selection is performed by analyzing the least values of the

32

P. Jayanthi and I. MuraliKrishna

Fig. 5 The prediction graph of India’s death cases

RMSE, MAPE, and MAE of the dataset. The model is statistically significant as the value of p is less than 0.05.

2.5 Graphs for the Four States: Confirm Cases and Death Cases The four states were considered based on the highest, average, and medium values of cases, respectively. Figure 7 shows the time series graph of CC of the four different states, namely AP, DE, MH, and TN. The highest case was found in MH with 5,71,495 and is known for the highest CV-19 prone state as of August 15, 2020, whereas TN has cases (3,26,520). DE curve seems to look better with the number of cases at hold and not much increase was found. The preventive measures increased with the more number of testing and treatment given. AP had fewer cases when compared to other states but during unlock down period the cases increased rapidly. Similarly, Fig. 8 shows the graph of death cases of AP, DE, MH, and TN. The death cases graph seems to increase consistently. But compared to other countries, India experienced few death cases initially and can maintain the trend, whereas in few countries one can see more deaths in March and April 2020.

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

33

Fig. 6 ACF and PACF graph – (a, b) actual death cases; (c, d) Day-wise increase in death cases; (e, f) predicted death cases of India

2.6 Prediction for Four States Using ARIMA Model The prediction of AP, DE, MH, and TN was modeled using AR and the outcome is shown in Table 4. The prediction for AP on September 15 was 5,05,168 and on October 15 was 6,13,208, and DE showed the values of 1,62,964 (September 15) and 1,40,909 (October 15). Similarly, the prediction for MH on September 15

34

P. Jayanthi and I. MuraliKrishna Table 3 Evaluation of tested ARIMA models

India’s Cases

Confirm

Death

Model ARIMA (2,0,3)

RMSE 3697.2446

MAE 340.0333

MAPE 15.6114

ARIMA (4,0,2)

3696.9228

339.795

15.5349

ARIMA (4,0,0)

3700.5294

330.829

15.6323

ARIMA(1,0,1)

124.1602

186.318

1.5829

ARIMA(2,0,2)

14.5123

7.5751

0.2469

Fig. 7 Time series graph of CC of four states as in Table 1

was 9,32,205 and on October 15 was 13,01,745; and TN values were 4,96,915 (September 15) and 6,24,097 (October 15). The prediction trend for Delhi looks better by decreasing whereas the other three states show an increasing trend in the graph. Figure 9 shows the AP prediction graph with a dashed line representing the prediction trend. The curve shows that the cases would decrease in the month of Nov 2020 the second week with 5,82,315. Figure 10 shows the y-axis yielding the range from −1.00 to +1.00 in increments of 0.50. The Delhi prediction graph is shown in Fig. 11 with the trend falling downward. The graph of AC and PAC of DE’s CC and predicted CC is shown in Fig. 12.

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

35

Fig. 8 Time series graph of death cases of four states as in Table 1 Table 4 The prediction for confirmed and death cases of India for the month of September, October, and November, 2020 Dates

Andhra Pradesh Con Cases

Prediction –AP CasesARIMA (2,0,4)

Delhi Con Cases

PredictionDE ARIMA (3,0,8)

Maharashtra Con Cases

Prediction -MH ARIMA (2,0,2)

Tamil Nadu Con Cases

15-Mar-20

1

9

33

1

15-Apr-20

473

1512

2337

1173

15-May-20

2205

8470

27524

9674

15-Jun-20

6163

41182

107958

44661

15-Jul-20

33019

15-Aug-20

273206

115346 271269

152134

267665 151998

571495

Prediction TNARIMA (2,0,0)

147324 568299

326320

326384

15-Sept-20

505168

162964

932205

496915

15-Oct-20

613208

140909

1301745

624097

15-Nov-20

582315

103363

1674853

697538

The prediction of MH is shown in Fig. 13 with the increasing trend up to the cases of 16,74,853 on November 15, 2020. Figure 14 shows the AC and PAC graph of CC and predicted CC in MH with the y-axis yielding the range from negative 1.00 to 1.00 in increments of 0.50. Similarly, the prediction of TN is shown in Fig. 15 with the increasing trend, and the graph of AC and PAC for CC and predicted CC in TN is shown in Fig. 16.

36

P. Jayanthi and I. MuraliKrishna

Fig. 9 The prediction graph of CC in AP

Fig. 10 The graph of AC and PAC (a, b). CC and predicted CC of AP (c, d)

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

Fig. 11 The prediction CC graph in DE

Fig. 12 The graph – (a, b) confirm and (c, d) predicted CC in DE

37

38

P. Jayanthi and I. MuraliKrishna

Fig. 13 The prediction graph of CC in MH

Fig. 14 The graph – (a, b) confirm and predicted CC of MH (c, d)

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

Fig. 15 The prediction graph of CC for TN

Fig. 16 The graph – (a, b) confirm and (c, d) predicted CC for TN

39

40

P. Jayanthi and I. MuraliKrishna

2.7 Tables of ARIMA Model for Four States Tables 5, 6, 7, and 8 are obtained using models of ARIMA (2,0,4) for AP, ARIMA (3,0,8) for DE, ARIMA(2,0,0) for TN, and ARIMA (2,0,2) for MH, respectively, and the models are better fit as p is less than 0 and proved with the computation of RMSE, MAE, and MAPE values as shown in Table 9. The least value of RMSE, MAE, and MAPE is the better model of AR. Table 10 is the summarized table showing p is equal to zero. The number of parameters required for each model with the order of difference and order of moving average. z-value is the value of the coefficient divided by the standard error. Hence, the models are statistically significant with a P-value less than the value of 0.05.

Table 5 ARIMA (2,0,4) model table for AP LL = 951.5739 AP Cons ARMAar L1 L2 ma L1 L2 L3 L4 /sigma

Coef. 211223.8 1.9995 −0.9957 −0.0348 −0.08170 0.1716 −0.3431 562.97

Prob > chi2 = 0.0000 SE z 221459.1 0.95 0.0081 244.27 0.0083 −119.58 0.0706 0.49 0.1221055 −0.67 0.0923176 1.86 0.0764535 −4.49 22.12 25.45

P 0.000 0.000 0.000 0.622 0.503 0.063 0.000 0.000

Wald chi2 (6) = 2.89e+06 95% CI −222827.9 645275.6 1.9793 2.0113 −1.012 −0.9794 −0.1037 0.1734 −0.3210 0.1576 −0.0092 0.3525 −0.4927 −0.1930 519.61 606.34

LL Log likelihood Table 6 ARIMA (3,0,8) model table for DE LL = −893.0357 DE Coef. Cons 75334.31 ARMAar L1. 1.0596 L2. 0.8537 L3 −0.9143 ma L1 20.8924 L2 48.7614 L3 8.7447 L4 35.3331 L5 −9.6181 L6 −63.7918 L7 23.5980 L8 35.8549 /sigma 4.3177

Prob > chi2 = 0.0000 SE z 44617.58 1.69 0.1806 5.87 0.3588 2.38 0.1795 −5.09 440.79 0.05 971.48 0.05 172.61 0.05 738.74 0.05 236.37 −0.04 1266.77 −0.05 504.28 0.05 694.26 0.05 86.53 0.05

P 0.000 0.000 0.017 0.000 0.962 0.960 0.960 0.962 0.968 0.960 0.963 0.959 0.480

Wald chi2 (6) = 1.34e+06 95% CI −12114.54 162783.2 0.70 1.41 0.15 1.56 −1.26 −0.56 −843.05 884.83 −1855.31 1952.83 −329.57 347.07 −1412.58 1483.25 −472.89 453.66 −2546.62 2419.03 −964.79 1011.99 −1324.87 1396.58 0 173.92

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

41

Table 7 ARIMA (2,0,0) model table for TN LL = −837.3072 Coef. TN 311123.4 Cons 1.9979 ARMAar L1. −0.9982 L2. 217.06 /sigma

Prob > chi2 SE 192710.8 0.0030 0.0029 9.3173

= 0.0000 z 1.61 665.40 −334.47 23.30

P 0.000 0.000 0.000 0.000

Wald chi2 (2) = 2.51e+07 95% CI −66582.86 688829.6 1.9920 2.0038 −1.0040 −0.9923 198.8027 235.3258

P 0.000 0.000 0.000 0.000 0.000 0.000

Wald chi2 (4) = 1.57e+08 95% CI −778719.2 3263625 1.998 2.001 −1.001 −0.998 −0.590 −0.324 −0.389 −0.126 732.964 846.242

Table 8 ARIMA (2,0,2) model table for MH LL = −992.081 MH Coef. Cons 1242453 1.9995 ARMAar L1 −0.9996 L2 −0.4576 ma L1 −0.2582 L2 /sigma 789.6033

Prob > chi2 SE 1031229 0.0007 0.0007 0.0680 0.6710 28.8978

= 0.0000 z 1.20 2558.88 −1359.51 −6.73 −3.85 27.32

Table 9 Comparison of four states ARIMA model

State-confirm

AP

DE

MH

TN

Model

RMSE

MAE

MAPE

ARIMA (2,0,2)

295.2297

5.4275

0.0293

ARIMA (2,0,4)

1506.0340

7639.282

14.0605

ARIMA (2,0,6)

3586.5272

7642.1556

14.0261

ARIMA(2,0,6)

6107.2590

3.1165

4.5412

ARIMA (3,0,6)

6106.9223

3.0752

4.5390

ARIMA (3,0,8)

6106.5931

2.9237

4.5402

ARIMA (2,0,1)

21837.7571

17097.31

16.6470

ARIMA (2,0,2)

21836.2431

17083.15

16.5853

ARIMA (3,0,2)

21840.6359

17084.37

16.5988

ARIMA (2,0,0)

13240.3471

9900.157

16.7730

ARIMA (2,0,2)

13241.7455

9901.461

16.8440

ARIMA (2,0,3)

13241.7390

9901.462

16.8442

42

P. Jayanthi and I. MuraliKrishna

Table 10 Summary of ARIMA model Country/state India, confirm India, death AP, confirm DE, confirm MH, confirm TN, confirm

Best model ARIMA(4,0,2) ARIMA(2,0,2) ARIMA(2,0,4) ARIMA(3,0,8) ARIMA(2,0,2) ARIMA(2,0,0)

Parameters AR (4)MA (2) AR (2)MA (2) AR (2)MA (4) AR (3)MA (8) AR (2)MA (2) AR (2)MA (0)

Coeff. 4279486 86418.99 211223.8 75334.31 1242453 311123.4

Standard error 3199113 81331.43 221459.1 44617.58 1031229 192710.8

z 1.34 1.06 0.95 1.69 1.20 1.61

p 0.000 0.000 0.000 0.000 0.000 0.000

2.8 Comparison of the Results with Other Researcher Results It is interesting to learn the results obtained from the other authors who modeled the AR model on the CV-19 dataset. According to Tandon H [14], the outcome for the AR model accuracy, linear trend, QL, S-Curve Trend, MA, and SE are the most accurate. The parameters of ARIMA (2,2,2) model were experimental that AR (2) and MA (2) parameters have a p-value of 0.000, 0.167, 0.000, and 0.000, respectively, which implies that the parameters are model significant. Maurya and Singh (2020) computed an AR model for the predicting techniques through Python and its libraries in the CV-19 dataset. The data was updated till June 26, 2020. It obtained the results with the lowest MSE (7240024855.066816) and the values of p (0) and q (0), and d (2) [15]. Roy (2020) [16] performed ARIMA (2,2,2) model and obtained RMSE (95.322) and MAE (50.109) with the dataset from January 30 to April 26, 2020 and from April 27, 2020 to May 11, 2020. Spatial distribution of disease risk analysis is carried out using the GIS platform. Results showed the low-risk zone in centraleast, north-east, and small northern areas in India; whereas the high-risk zone in the west, south, south-west, and central-north districts of India. According to the AR model (Chaurasia and Pal 2020), the number of death cases will increase as predicted with the dataset as of May 29, 2020 [17]. Verma (2020) [18] modeled ARIMA (2,1,2) that predicts better the case of CV-19 deaths in India with RMSE (17.77). But, in the case of CV-19 deaths in MH, the FTS model is the best-fitted model for forecasting with a lower RMSE (5.16) value. Papastefanopoulos (2020) modeled AR whose performance in terms of RMSE for ten countries with the most confirmed CV-19 cases as of May 4, 2020, was obtained as follows: the USA (0.007421), Spain (0.080094), Italy (0.005628), the UK (0.005484), France (0.060824), Germany (0.006431), Russia (0.001536), Turkey (0.004442), Brazil (0.004194), and Iran (0.002628) [19]. Hernandez-Matamoros (2020) [20] proposed an algorithm that would calculate the AR model for 145 countries. The results showed that more models would be created to predict the CV-19 behavior using a few variables such as humidity, climate, and culture among others. RMSE value for each region was noted and a few of them are listed below. North America region has 13 countries and obtained an average RMSE (640.61). The USA presents the bigger RMSE among the 145

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

43

countries (7749.99), this country has the largest population in the region (329.06 pp Mp). On the other hand, Spain presents an RMSE (1892.33) with 46.73 pp Mp, Italy presents an RMSE (566.88) with 60.55 pp Mp, and the UK presents an RMSE (728.13) with a 467.53 pp Mp, Germany has 83.51 pp Mp and presents an RMSE (1075.02), and Russia presents an RMSE (958.44) with a 145.87 pp Mp. Brazil presents an RMSE (591); this country has the largest population in the region (211.04 pp Mp). In Paraguay presents the lowest RMSE. Brazil has almost 30 times the population of Paraguay (7.05 pp Mp). The South American region produced the average RMSE (104.78). But, the Asia region included 40 countries, where Turkey presents the slightly bigger RMSE (696.35) with a population (83.43 pp Mp). But the RMSE for China (117.88) and India (250.83), and RMSE for Egypt (84.08) were obtained. In contrast, Namibia produced an RMSE close to zero. The countries from African region obtained the RMSE average (13.8) whereas the Oceania region includes four countries. Australia obtained RMSE (24.76) with a population (25.20 pp Mp). Fiji had the lower RMSE with a population of less than 1 pp Mp. Lee (2020) proposed ARIMA modeled that obtained a forecast waveform with an average error rate (2.84%) [21]. Masum (2020) modeled the ARIMA (2, 3) model which was selected as the best model based on the lowest MAPE (14.1). The minimum MAPE, RMSE, and MAE for the first, second, third, fourth, and fifth rounds of forecasts were achieved by observing 2, 3, 2, 9, and 14 lag days, respectively. Best forecasting accuracy (MAPE score: 0.12; RMSE: 0.23 × 104, MAE: 0.19 × 104) was achieved for the fifth round of forecast where the number of training data was maximum compared to other rounds. Masum performed 15 × 5 × 30 = 2250 experiments in total to understand the best model based on MAPE, RMSE, and MAE [22]. Sulasikin (2020) proposed the best AR model prediction compared to the Holts-Winters’ and Holt’s method. ARIMA model has an excellent fit with R2 (0.969) and MSE (43,643,545) and RMSE (208,910) [23]. Kumar and Susan (2020) proposed an ARIMA model that performed a better fit model. The results obtained were the best MAPE scores for US (0.586) and UK (1.481) data by AR and FBProphet, respectively. From the results obtained, it was clear that AR has far better performance as compared to the FBProphet model with respect to all types of error measures, i.e., MAE, RMSE, RRSE, and MAPE [24]. Yang (2020) proposed ARIMA(0,2,0) in the confirmed new cases, the ACF and PACF diagram of the second-order difference sequence showed that the correlation values did not exceed the significant boundary (0.5) [25]. Tran (2020) [26] focused on the forecasting of the pandemic spread of SARSCoV-2 in Iran. The summarized results for forecasting are (i) Based on AIC and the Ljung–Box-Q statistic test, the ARIMA models are probably adequate for data of the TCC (AR (0,2,0), TCNC (AR (0,1,0), TD (AR (1,1,0), TND (AR (0,1,0), GRCC (AR (2,1,1), and GRD (AR (1,0,0); (ii) Iran may be successful in haltering the SARS-CoV-2 pandemic because the ARIMA model forecasts a stable in all variables in the near future. Kumari (2020) proposed the AR model that would predict a number of cases that

44

P. Jayanthi and I. MuraliKrishna

shows a good agreement with 0.9992 R-squared scores to the actual values [27]. Vakula (2020) implemented the Sigmoid, SEIR, AR and LSTM models. On comparing the two forecasting plots, it is found that the AR model has predicted the numbers close to the actual numbers [28]. Istaiteh’s (2020) AR model provided promising results with MAPE (14.14) and RMSLE (0.17) in predicting accumulated cases, especially in the countries that have linear cases curve. But it was observed that it did not perform well with the non-linear curves, and the drawback of this model was that every country needs a separate AR model which is computationally expensive and time-consuming [29]. Arun (2021) performed two models, namely, GRU and LSTM, based on GRU models for three countries, namely, Peru (≈26,000), Chile (≈12,100), and the UK (≈49,000). However, the LSTM model of the following countries predicted the values of ≈31,000, ≈12,100, and ≈49,000 cumulative death cases by the end of September 2020. Whereas for fatalities- RMSE using the LSTM model in other countries, namely, Brazil (3.66E+04), India (2.57E+02), Russia (7.20E+01), South Africa (8.74E+02), Mexico (3.04E+02), and Iran (5.20E+01) were obtained [30]. Malki (2021) performed the following models, namely, AR, SARIMA, ML (RF), and DL model (LSTM). The best predicting model is SARIMA whose parameters are selected based on the lowest values of AIC, and P-values that are less than 0.05. The following SARIMA (9,0,8) × (0,0,0,3) model has the lowest AIC values (−2199.02) [31]. Ganiny (2021) [32] proposed that ARIMA (7,2,2) models were good fitting and predicted capabilities for all three categories of datasets. The predicted values of ARIMA (7,2,2) models were very close to the actual values. The cumulative diagnosed cases would escalate by 63.54%, which is lesser than the current value of 65% and therefore would be promising. The predictions made as of August 24, 2020, the expected number of cumulative detected cases would increase nearly to 3,800,989, recoveries (2,110,697), and the cumulative number of deaths (56,150). India’s cases were predicted to exceed two million cases mark as of August 5, 2020, and three million cases as of August 17, 2020, and it was true. Chaurasia and Pal (2021) [33] showed the estimated mortality rate by the AR model and the regression model on CV-19 dataset. For mortality rate, the correlation coefficients of attributes (MAE, MSE, RMSE, and MAPE) were used, where the average absolute percentage error (99.09%). Sharma and Nigam (2020) found ARIMA (5,2,5) model to be the best-fitting model for CV-19 cases in India with RSE (4178) and MAPE (0.85%) [34]. The authors [35] showed ARIMA (1, 0, 17) model with P = 0.009 to be the best model for the Telangana state (India).

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

45

2.9 Spatial Distribution of Predicted COVID-19 Confirm Data in India India stands third in the world followed by the USA and Brazil with the number of total CC breaking the mark of 100,000 on May 19, 200,000 cases on June 3, 2020 and 1,000,000 on July 17, 2020, respectively. Though India’s disease rate is among the lowest in the world at 2.41% as of July 23, 2020, and is steadily declining. It has an impact on education, transport, economy, and commercial, and CV-19 lockdown in India has left tons of millions of migrant workers unemployed. The prediction for September 15, October 15, and November 15, 2020, is shown in Fig. 17. It clearly shows where and how many cases would be on the coming September 15, 2020. These geospatial distributions of the predicted confirm cases for all the states are carried out using ArcGIS software. India is assumed to have native immunity because of which the number of cases is comparatively less than the abroad countries or developed countries like Italy, the USA, or the UK and death rate is comparatively lower than other countries of the UK and the USA. Thus, the geomap helps to understand the data visualization in a state or country. Despite all these cases the country tries to have a positive look toward improving the economic growth the country. Great efforts were imparted by the health care ministries and police department in controlling the movement in March and April, 2020. It is the responsibility of an individual citizen to avoid the spread of the Covid or not.

2.10 Temporal Distribution of Confirm Cases: Day-Wise Trends in Each State The highest confirmed case (47,989) was found in East Godavari in AP; similarly, the worst affected district in MH is Pune with 1,47,671 cases; Chennai in TN has the maximum number of cases (1,24,071) as of August 26, 2020. Figure 18 shows the day-wise increase for all the different states; the graphs show AP, TS, MH, and UP have an increasing trend in cases.

3 State District Distribution of COVID-19 Cases in AP AP is the second top state in the list of CV-19 cases in India. As of August 30, 2020, AP was leading with the second highest cases with 4,24,767 after MH (7,80,689) followed by TN (4,22,085) and DE (1,73,390). In AP, the districts, namely, East Godavari followed by Kurnool district and Anantapur (Fig. 19) are the major cities that are affected by CV-19. However, the virus has spread to all the states except Lakshadweep and all the districts of the states too.

46

P. Jayanthi and I. MuraliKrishna

Fig. 17 Spatial distribution of predicted CC in India – September 15, October 15, and November 15, 2020

4 Interpolation and Choropleth Map of AP Districts The areal interpolation is implemented for AP district CC and the spatial distribution of CC is shown in the choropleth map in Fig. 20. Interpolation predicts the district cases can be seen clearly in the figure with the color deep blue and K-Bessel curve

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

Fig. 18 Day-wise increase in CC in each state of India

47

48

Fig. 18 (continued)

P. Jayanthi and I. MuraliKrishna

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

Fig. 18 (continued)

49

50

Fig. 18 (continued)

P. Jayanthi and I. MuraliKrishna

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

51

Fig. 18 (continued)

Fig. 19 CC in each district of AP state

model flattens at the distance of 0.889 on the x-axis is a better fit for interpolation. Spherical starts at 5.089 on y-axis and flattens at 0.60 on x-axis.

5 Lockdown Period Graphs in Six States of India The lockdown in a few states was declared based on the rise of the cases individually. The single-day CCs increase in MH was 68,631 cases (April 18, 2021). The curve (Fig. 21) shows a drop in cases during the lockdown whereas, in AP and KA, the

52

P. Jayanthi and I. MuraliKrishna

Fig. 20 Areal interpolation of AP district CC and choropleth map of AP

cases were raising in spite of the lockdown. DE, TS, and UP show a clear drop in cases and a trend of curve fall. Major research is still going to understand the rise in AP and KA.

6 IoT for COVID-19 Healthcare Karanika et al. (2020) [36] proposed the use of interpretable ML to deliver the features significant for any data processing activity. The authors suggested that the quality of data plays a significant issue in any application of data analytics to support decision-making. In this COVID study, IoT devices can interact to exchange and process COVID-19 data analysis and monitor processing activities using IoT edge computing (EC) in Fig. 22. IoT devices are therefore connected to EC nodes to report the data collection and secure the data quality at the network edge. Multiple services and applications are executed over the massive data volumes collected by the IoT devices. But the decision of the researcher community is whether the data be moved to the cloud for further processing or not. The final decision process is carried out at the edge of the network due to network bandwidth, latency, and other concerns about data privacy. It was observed that EC involves heterogeneous nodes close to IoT devices and end users perform various activities and deliver statistical analytics over the collected data.

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

53

Fig. 21 Lockdown graphs in few states

7 Conclusion ARIMA model was found to be the best suitable interpretable model in analyzing the CV-19 data. ARIMA (4,0,2) was best for India confirm cases with z (1.34); ARIMA (2,0,2) for India death cases with z (1.06); similarly, ARIMA (2,0,4) for Andhra Pradesh; ARIMA (3,0,8) for Delhi cases. Maharashtra model is ARIMA (2,0,2) and ARIMA (2,0,0) for Tamil Nadu. This model can help the health and police department to understand the growing trend of the cases. Despite the lockdown in a few states, the trend of cases increased in KA and UP. The private and government departments would thus take precautions and preventive measures from increasing

54

P. Jayanthi and I. MuraliKrishna

Fig. 22 IoT edge computing for COVID data analysis

the cases. The model helps to make necessary arrangements for treating the citizens of the country and overcoming the pandemic in the future.

References 1. Maleki, M., Mahmoudi, M. R., Heydari, M. H., & Pho, K. (2020). Modeling and forecasting the spread and death rate of coronavirus (COVID-19) in the world using time series models. Chaos, Solitons and Fractals, 140, 110151. 2. Duan, X., & Zhang, X. (2020). ARIMA modelling and forecasting of irregularly patterned COVID-19 outbreaks using Japanese and South Korean data. Data in Brief, 31(105779), 1–6. 3. Khan, F. M., & Gupta, R. (2020). ARIMA and NAR based prediction model for time series analysis of COVID-19 cases in India. Journal of Safety Science and Resilience, 1, 12–18. 4. Ceylan, Z. (2020). Estimation of COVID-19 prevalence in Italy, Spain, and France. Science of the Total Environment, 729(138817), 1–7. 5. Sahai, A. K., Rath, N., Sood, V., & Singh, M. P. (2020). ARIMA modelling & forecasting of COVID-19 in top five affected countries. Diabetes and Metabolic Syndrome: Clinical Research and Reviews, 14, 1419–1427. 6. Benvenuto, D., Giovanetti, M., Vassallo, L., Angeletti, S., & Ciccozzi, M. (2020). Application of the ARIMA model on the COVID- 2019 epidemic dataset. Data in Brief, 29(105340), 1–4. 7. Feroze, N. (2020). Forecasting the patterns of COVID-19 and causal impacts of lockdown in top five affected countries using Bayesian Structural Time Series Models. Chaos, Solitons and Fractals, 140(110196), 1–8. 8. Singh, S. (2020). Data on social and health vulnerability in rural India: A case of covid-19. Data in Brief, 31(106020), 1–4. 9. Lai, Y., & Dzombak, D. A. (2020). Use of the autoregressive integrated moving average (ARIMA) model to forecast near-term regional temperature and precipitation. Weather and Forecasting, 35, 959–976. https://doi.org/10.1175/WAF-D-19-0158.1

ARIMA and Predicted Geospatial Distribution of COVID-19 in India

55

10. Iyyanki, M. K., & Prisilla, J. (2020). The prediction analysis of Covid-19 cases using ARIMA and KALMAN filter models: A case of comparative study. Internet of Medical Things for Smart Healthcare, Studies in Big Data, 80, 167–191. 11. Harder, H. (2020). Interpretable-machine-learning-models. Available at https:// towardsdatascience.com/ 12. Chatterjee, S. (2020). ARIMA/SARIMA vs LSTM with ensemble learning insights for time series data. Available at https://www.datasciencecentral.com/ 13. Prabhakaran, S. (2021). ARIMA model – Complete guide to time series forecasting in Python. Available at https://www.machinelearningplus.com/ 14. Tandon, H., Ranjan, P., Chakraborty, T., & Suhag, V. (2020). Coronavirus (COVID-19): ARIMA based time-series analysis to forecast near future. arXiv:2004.07859, 1–11. 15. Maurya, S., & Singh, S. (2020). Time series analysis of the Covid-19 datasets. In IEEE international conference for innovation in technology (pp. 1–6). IEEE. 16. Roy, S., Bhunia, G. S., & Shit, P. K. (2021). Spatial prediction of COVID-19 epidemic using ARIMA techniques in India. Modeling Earth Systems and Environment, 7(2), 1385–1391. 17. Chaurasia, V., & Pal, S. (2020). Application of machine learning time series analysis for prediction COVID-19 pandemic. Research on Biomedical Engineering, 38, 35–47. 18. Verma, P., Khetan, M., Dwivedi, S., & Dixit, S. (2020). Forecasting the covid-19 outbreak: An application of ARIMA and fuzzy time series models, pp. 1–15. 19. Papastefanopoulos, V., Linardatos, P., & Kotsiantis, S. (2020). COVID-19: A comparison of time series methods to forecast percentage of active cases per population. Applied Sciences, 10(3880), 1–15. 20. Hernandez-Matamoros, A., Fujita, H., Toshitaka, H., Hayashi, T., & Perez-Meana, H. (2020). Forecasting of COVID19 per regions using ARIMA models and polynomial functions. Applied Soft Computing, 96(106610), 1–14. 21. Lee, D., Lee, D., Choi, M., & Lee, J. (2020). Prediction of network throughput using ARIMA. In International conference on artificial intelligence in information and communication (ICAIIC) (pp. 1–5). IEEE. 22. Masum, M., Shahriar, H., Haddad, H. M., & Alam, M. S. (2020). r-LSTM: Time series forecasting for COVID-19 confirmed cases with LSTM-based framework. In IEEE international conference on big data (Big Data) (pp. 1374–1379). IEEE. 23. Sulasikin, A., Nugraha, Y., Kanggrawan, J., & Suherman, A. L. (2020). Forecasting for a datadriven policy using time series methods in handling COVID-19 pandemic in Jakarta. In IEEE international smart cities conference (ISC2), Piscataway, NJ, USA, pp. 1–6. 24. Kumar, N., & Susan, S. (2020). COVID-19 pandemic prediction using time series forecasting models. In 11th international conference on computing, communication and networking technologies (ICCCNT) (pp. 1–7). IEEE. 25. Yang, Q., Wang, J., Ma, H., & Wang, X. (2020). Research on COVID-19 based on ARIMA model—Taking Hubei, China as an example to see the epidemic in Italy. Journal of Infection and Public Health, 13, 1415–1418. 26. Tran, T. T., Pham, L. T., & Ngo, Q. X. (2020). Forecasting epidemic spread of SARS-CoV2 using ARIMA model (case study: Iran). Global Journal of Environmental Science and Management, 6, 1–10. 27. Kumari, R., Kumar, S., Paonia, R. C., Singh, V., Raja, L., Bhatnagar, V., & Agarwal, P. (2021). Analysis and predictions of spread, recovery, and death caused by COVID-19 in India. Big Data Mining and Analytics, 4(2), 65–75. 28. Vakula, R. J., & Jakka, A. (2020). Forecasting COVID-19 cases in India using machine learning models. In International conference on smart technologies in computing, electrical and electronics (pp. 466–471). IEEE. 29. Istaiteh, O., Owais, T., Al-Madi, N., & Abu-Soud, S. (2020). Machine learning approaches for COVID-19 forecasting. In International conference on intelligent data science technologies and applications (IDSTA) (pp. 50–57). IEEE.

56

P. Jayanthi and I. MuraliKrishna

30. Arun, S. K. E. K., Kalaga, D. V., Kumar, M. S., Kawaji, M., & Brenza, T. M. (2021). Forecasting of COVID-19 using deep layer recurrent neural networks (RNNs) with gated recurrent units (GRUs) and long short-term memory (LSTM) cell. Chaos, Solitons and Fractals, 146(110861), 1–12. 31. Malki, Z., Atlam, E. S., Ewis, A., et al. (2021). ARIMA models for predicting the end of COVID-19 pandemic and the risk of second rebound. Neural Computing and Applications, 33, 2929–2948. 32. Ganiny, S., & Nisar, O. (2021). Mathematical modeling and a month ahead forecast of the coronavirus disease 2019 (COVID-19) pandemic: An Indian scenario. Modeling Earth Systems and Environment, 7, 29–40. 33. Chaurasia, V., & Pal, S. (2020). COVID-19 pandemic: ARIMA and regression model-based worldwide death cases predictions. SN Computer Science, 288, 1–12. 34. Sharma, V. K., & Nigam, U. (2020). Modeling and forecasting of COVID-19 growth curve in India. Transactions of the Indian National Academy of Engineering, 5, 697–710. 35. Jayanthi, P., & Muralikrishna, I. (2021). ARIMA Model to Predict the Covid-19 Pandemic Cases in Telangana State. 2021 IEEE International India Geoscience and Remote Sensing Symposium (InGARSS), Ahmedabad, India, 2021, pp. 185–189 36. Karanika, A., Oikonomou, P., Kolomvatsos, K., & Anagnostopoulos, C. (2020). On the use of interpretable machine learning for the management of data quality. https://arxiv.org/abs/ 2007.14677

Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in Healthcare IoT Systems Kevser Sahinbas and Ferhat Ozgur Catak

1 Introduction Cloud computing technology, one of the rapidly developing technologies in the information technologies, provides secure and location-independent storage, computing, and various application services. Besides, with the production of new generation hardware with different sensors, low cost, and efficient energy consumption, the Internet of Things that transforms any sensitive data into network data by wireless sensor networks has been widely used in the detection, monitoring, control, and smart management system [4, 7, 16]. New generation applications that benefit from these two technologies provide important services in different fields. For example, the advantages of the IoT and cloud computing technologies have been extensively adopted in healthcare, such as remote patient care and hand hygiene monitoring systems [12, 27, 37]. Health devices: Information about the devices such as their instant location, calibration status, and measures against theft can be followed instantly. In addition, significant progress has been made in monitoring the health of people outside of the hospital. People in their homes, workplaces, or anywhere can communicate with health service providers and organizations with IoT technology and access healthcare services remotely. With the development of 5G technology, it is thought that IoT will take a more active role in our life. IoT technology is becoming the most powerful and flexible technology on the market. Therefore, the market is currently flooded with various devices and systems in multiple fields of technology development and commercialization. With the

K. Sahinbas Istanbul Medipol University, Istanbul, Turkey e-mail: [email protected] F. O. Catak () University of Stavanger, Stavanger, Norway e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_3

57

58

K. Sahinbas and F. O. Catak

increasing number of such products and services, consumers are starting to perceive the value of privacy and security of their personal data in new ways [5]. In addition to privacy concerns, consumers are concerned about the growing usage of Internetbased devices and the resulting surveillance in public regarding personal data. There is increasing acceptance of data protection as a fundamental right in developed countries, and there is growing public awareness regarding personal data privacy. Therefore, the market is highly concentrated in the area of data protection and privacy with IoT data. Real-time and accurate data collection from sensors is one of the critical functions in the Internet of Things. Generally, in sensor-based systems, an asset’s environment, structure, or various properties are detected by sensors and sent to a base station and from there to a data processing center over WiFi, a wired communication infrastructure, or multi-hop connections. After the data reach the processing center, intelligent decisions can be made based on the perceived information or the quality of the services provided. Figure 1 indicates the general system architecture of IoT systems in healthcare [30]. The number of devices connected to the Internet has been increasing recently, and these devices are in homes and workplaces with the security risks they contain. Connecting billions of devices, IoT encompasses various things from wearables, smart home appliances, vehicle sensors, and medical devices. With the ubiquity of IoT technology, it is also exposed to many cyberattacks. Those who take control of the devices can dangerously use these devices. This situation requires the use

Healthcare Applications information integration and management

Data analysis analyst

Gateway of IOT

Access network

Application Layer

Wireless Sensor Networks

Fig. 1 The architecture of healthcare IoT systems

Network Layer

Perception Layer

Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in. . .

59

of security systems in IoT technology. Thus, privacy and security issues must be considered and addressed. To meet such requirements, we will propose a model that overcomes privacy problems by applying federated learning (FL). We summarized some of the surveyed researches in the literature in relation to the privacy-preserving healthcare IoT data. Gou et al. [7] present a mutual privacypreserving k-means strategy (M-PPKS) model that applies homomorphic encryption to handle privacy issues. Sage [17] provides a privacy-preserving outline by implementing privacy-sensitive data transmission. It ensures content confidentiality by using encryption methods in its work. It provides the anonymity of the receiver using the broadcast strategy that allows hiding the potential doctor. Huang et al. [9] provide a privacy plan for e-health services by tracking the patients’ physical condition to obligate anonymous authentication. By encrypting data exchanged using elliptic curve cryptography (ECC), their plans also provide content-driven privacy. The study ensures unlinkability between the identity and the biometric information of each patient. Lu et al. [20] provide an opportunistic computing framework that maintains secure privacy for the m-Healthcare emergency. The purpose of the study is to aid patients monitor and protect personal health data in pervasive environments. It provides content privacy based on cryptographic primitives. Zhang et al. [39] proposed a D2Dassist data transmission protocol (LSD) that helps integrity, data confidentiality, and mutual authentication through applying cryptographic primitives. However, other contextual privacy requirements have not been met as advertised. Yang et al. [35] structured a privacy-preserving e-health system, which is a fusion of IoT, big data, and cloud storage. Yang et al. [35] have configured a privacy-preserving e-health system that is a fusion of big data and cloud storage. They provide access control as well as content privacy based on attribute-based encryption. The main benefits of federated learning are: • FL allows devices such as smartphones to learn a shared prediction model cooperatively while keeping the training data on the device rather than uploading and storing it on a central server. It also protects privacy because devices will be unable to access or save the data shared with them. As described in the paper by Bonawitz et al. [3], FL can facilitate data sharing with multiple entities on the network to improve data security and availability in the network . • Moves model training to the edge, incorporating devices such as smartphones, tablets, and the Internet of Things and institutions such as hospitals that must adhere to strict privacy rules. Keeping personal data local has a significant security benefit. • Real-time prediction is possible since prediction tasks happen on the device itself. FL reduces the time delay created by transmitting raw data to a central server and then delivering the results back to the device. It also improves privacy because the training is done locally rather than centrally in one location. To increase model performance, the model is then shared with numerous peers and dispersed over the network.

60

K. Sahinbas and F. O. Catak

• The prediction procedure works even with no Internet connection since the models are stored on the device. • FL reduces the amount of hardware infrastructure required. FL models require very little technology, and what is accessible on mobile devices is more than adequate. The road map of the chapter is organized as follows. we explain federated learning in Sect. 2. Section 3 contains dataset and experiment. In Sect. 4, we present an extensive privacy analysis and performance evaluation. Section 5 shows threats to validity. Lastly, we conclude the chapter at Sect. 6.

2 Federated Learning Types Federated learning (FL) broadly can be categorized into three types, such as horizontal, vertical, and hybrid [35] based on the data partition. FL is immensely helpful for building models where data are shared across different domains. The hybrid FL is based on transfer learning, while horizontal FL leverages the data parallelism and the vertical FL leverages model parallelism [26]. In the following sections, we will provide different federate learning and aggregation methods.

2.1 Horizontal Federated Learning Horizontal FL, which is mentioned to homogeneous FL [8], is defined as the circumstances in which datasets on the devices share the same features space but are different in instances. Basically, horizontal FL splits the datasets horizontally and then subtracts the part of the data where the user attributes are the same. Still, the users are not exactly the same for training. In this way, the user sample size can be increased by using horizontal FL. In horizontal FL, as shown in Fig. 2, all parties calculate and load local gradients, and then the central server aggregates to provide a global model. Homomorphic encryption [1], secure aggregation [6], and differential privacy [24] can provide security of the transferring process of gradients in horizontal FL. Figure 2 shows a typical horizontal FL overview.

2.2 Vertical Federated Learning Vertical FL [34] is referred as heterogeneous FL [36], in which users’ training data share the same sample space but have different feature spaces.

Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in. . .

Central Server

Global Model

61

1

1

2

2 1

1 2

2

1

1

2

2

Fig. 2 Horizontal FL overview

Collaborator C 4 3

3

4

1

1

2 Data from A

2

Data from B

Fig. 3 Vertical FL overview

Vertical FL involves dividing datasets vertically and removing the part of the data for training where the users are the same but the user characteristics are different. As a result, vertical FL has the potential to enhance the feature dimension of training data. In practice, several common solutions for vertical division of data concerns have been implemented in the literature, such as safe linear regression, classification, and gradient descent [31]. Figure 3 shows a typical vertical FL overview. In Fig. 3, arrow 1 refers sending public key, 2 is exchanging encrypted value, 3 is sending encrypted result, and 4 is updating the model.

2.3 Aggregation Algorithms Different aggregation techniques that compound the local model updates from all the clients participating in the training cycle have been presented in this area [25].

62

K. Sahinbas and F. O. Catak

Aggregation algorithms may be divided into three types: centralized, hierarchical, and decentralized [15].

2.3.1

Centralized Aggregation

In the centralized aggregation configuration, a single parameter server provides a global model that calculates the average gradients gathered from clients and then coordinates their updates using a centralized aggregation algorithm. Federated averaging changes in each implementation of FL in terms of the pre-config parameters. We have presented primarily used centralized aggregation algorithms. Federated averaging (FedAvg) is a primitive and most extensively applied federated learning algorithm introduced in Google’s implementation of FL, based on stochastic gradient descent (SGD) optimization algorithm. FedAvg includes a central server acting as a coordinator that stores the overall predictive model transmitted to a randomly selected subset of devices. Devices in the federation train the model using its local data and are responsible for sending parameter updates back to the server. Lastly, the server aggregates these updates and creates a new global model [24]. The other algorithm is the federated stochastic block coordinate descent (FedBCD) algorithm that has a similar principle with FedAvg and targets total rounds of communications, skipping updates for each iteration to achieve desired accuracy rate [19]. FedProx is a modified version of the FedAvg algorithm presented in [14] to overcome heterogeneity in FL and allows multiple iterations on each compute resource while minimizing a cost function based on a local loss function and a global model. FedMA constructs a shared model for CNNs and LSTM-based ML model updates in FL environments [32]. FedMA averages models on the central server by mapping and averaging hidden elements in neural networks, such as neurons and channels, on a layer-by-layer basis. Stochastic controlled averaging for FL (Scaffold) [64] reduces communication rounds by using stateful variables in distributed computing resources [11]. Attentive federated aggregation (FedAttOpt) adds an enhanced mechanism for modeling aggregation on the central server of the computed FL [10].

2.4 Hierarchical Aggregation A hierarchical architecture is utilized by using multiple parameter servers that named global parameter server (GPS) and multiple region parameter servers (RPS) to decrease models’ transfer time between a parameter server and computing resources [38]. Each RPS is applied in a cell base station to which computing resources can be connected with low latency. For the hierarchical approach, various

Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in. . .

63

algorithms are proposed. Hierarchical federated learning (HFL) [38], HierFAVG [18], LanFL [38], and HFEL [21] are applied to perform model aggregation.

2.5 Decentralized Aggregation Dependency on the central server is excluded to enable the model collection, and peer-2-peer topology is followed for communication in a decentralized approach [29]. Any global model is not seen, and each computing resource develops its model to share information with the other neighbors. For decentralized aggregation, various algorithms are proposed. In the decentralized SGD model (D-SGD), each computing resource keeps a global model’s local copy and uses its neighbors’ models to update the local copy. Average with-communication (AWC) [13] and average-before-communication (ABC) [33] are two common types of execution of D-SGD.

3 Dataset In this chapter, we used a publicly available Mobile Health (MHEALTH) dataset [2] of UCI Machine Learning Repository.1 The MHEALTH (Mobile HEALTH) dataset contains recordings of body motion, and vital signs were taken, while 10 volunteers of various profiles engaged in multiple physical activities. Sensors are installed on the right wrist, subject’s chest, and left ankle to track the movements of different body components, including acceleration, rate of rotation, and magnetic field orientation. The sensor on the chest can also take 2-lead electrocardiogram (ECG) readings, which may be utilized for basic cardiac monitoring, testing for different arrhythmias, or examining the effects of exercise on the ECG. Figure 4 shows the column values’ histogram distribution. According to the figure, almost all columns have a normal distribution.

4 Evaluation Section 4 covers the findings about experiments about DI model for IoT devices. In Sect. 4.2, we present the experiment design, followed by experiment execution in Sect. 4.3, and results in Sect. 4.5. Section 5 provides the threats to validity of our experiment.

1 http://archive.ics.uci.edu/ml/datasets/mhealth+dataset.

64

Fig. 4 Columns histogram values

K. Sahinbas and F. O. Catak

Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in. . .

65

4.1 System Overview The main aim of this chapter is to build a DL model for different hospitals based on their corresponding patient records. The proposed federated-learningbased approach mainly consists of two phases: (Phase I) the model building at individual hospital phase and (Phase II) the model combination at a trusted authority (Fig. 5). As shown in Fig. 5, Phase I is responsible for data pre-processing and model building with input dataset, .D, in each hospital. A standard DL training is performed at this phase, and categorical cross-entropy loss functions are used in weight optimization with the exact same DL architecture (i.e., the number of neurons at each layer, the number of layers, optimization method, and activation functions). The primary outcome in each hospital is their own model, which is the model with the best prediction accuracy and lowest loss over the entire patient records. In Phase II, the trusted authority collects all models from all hospitals. Figure 5 illustrates multiple models with n different hospitals (Hospital-1, Hospital-2, .. . . Hospital-n), which are generated from their private datasets. The trusted authority is responsible for combining all layers and neurons (i.e., mean value of each weight).

Fig. 5 System overview

66

K. Sahinbas and F. O. Catak

Then the trusted authority sends the combined model to each hospital for the retraining process. This process continues until there is no change in model weights. After this stage, the combined model is given to all hospitals and used as a prediction model. Algorithm 1 shows the pseudocode implementation of the overall process.

Algorithm 1: Centralized merging of privacy-preserving model Input: T is the number of clients in FL, Dt = {(Xt , yt )|Xt ∈ Rm×n , yt ∈ Rm } is the local dataset at client t; L is the number of layers in the model h, k is the number of iteration. Output: The final privacy-preserving model h // Randomly initialize the final model’s weights 1 h ← rand(w) // Send the randomly initialized model to each client 2 foreach t ∈ T do 3 send_model_to_client (h, t) // Iterative training 4 for i = 0; i < k; i + + do // Models training stage at each client // At each client t, perform a model building, ht , using its local dataset, Dt 5 foreach t ∈ T do // Train the local model 6 ht ← train(Xt , yt ) // Send the trained model to the central authority 7 send_model(ht )

8

9 10 11

// Models merging stage // Iterate each layer in the models foreach l ∈ L do // Calculate the mean value of the layer t weight vector of each model. h(t) ← mean(h(l) t ) // Send the merged model to each client foreach t ∈ T do send_model_to_client (h, t)

12 return h

4.2 Experiment Design We designed a series of experiments by uniformly varying 5 different number of devices, .t ∈ [3, 5, 10, 15, 30]. Each experiment was executed 30 times, results of which, corresponding to each t, are averaged to smooth the plotting. We selected the

Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in. . .

67

Fig. 6 DL model architecture in each IoT device

best hyper-parameters for the DL models for the dataset we used in the experiments using a simple grid search. It turned out that the best hyper-parameters are the RMSprop optimization with a learning rate of 0.01. Figure 6 shows DL models for each dataset.

4.2.1

Research Questions

We aim to explore privacy-preserving of medical IoT devices and explore federated learning to protect privacy. We formed the following research questions (RQs) and designed the experiments to answer them: • RQ1: Is federated-learning-based privacy protection applicable for medical IoT devices? Studying this RQ is necessary as a positive answer to it motivates us to further investigate whether federated learning can help to protect the privacy. • RQ2: Is there any relationship between prediction performance and the number of devices while protecting the privacy of medical IoT devices? RQ2 helps us to assess the quality of a DL model’s prediction while protecting the privacy of the dataset.

4.3 Experiment Execution All the experiments were performed using Python scripts and ML libraries: Keras, TensorFlow, and Scikit-learn, on the following machine: 2.8 GHz Quad-Core Intel Core i7 with 16GB of RAM. The dataset has been divided into two parts: 20% test dataset and 80% training dataset. In the training dataset, we randomly distributed subsets of the dataset to the simulated IoT devices and used them to build the DL model. The number of epochs was set to 50, and the epoch size was fixed for all the models in each IoT device.

68

K. Sahinbas and F. O. Catak

4.4 Prediction Performance Metrics In this chapter, we used four metrics (overall prediction accuracy, recall, precision, and .F1 score) that are common measurement metrics to evaluate classification accuracy and to find an optimal classifier in machine learning [22, 23, 28]. Precision is defined as the fraction of retrieved samples that are relevant: P recision =

.

Correct . Correct + F alse

(1)

Recall is defined as the fraction of relevant samples that are retrieved: Recall =

.

Correct . Correct + Missed

(2)

The .F1 score is defined as the harmonic mean of precision and recall: F1 = 2 ×

.

P recision × Recall . P recision + Recall

(3)

4.5 Results 4.5.1

Results for RQ1

To answer this RQ, first, we obtain values of the prediction performance metrics (i.e., accuracy, precision, recall, and .F1 ) of the DL models, which are shown in Fig. 5. Figure 7 delivers the prediction performance for the dataset in terms of accuracy, precision, recall, and .F1 . As shown in the figure, there is a negative relationship between the number of clients in the federated learning and prediction accuracy. With the number of clients, a decrease is observed in the final model classification performance. Thus, as shown in the figure, the DL model performance drops considerably as the number of clients is increased. Moreover, although the total number of clients varies, the final model classification accuracy is almost equal to the base model’s accuracy. The negative relationship in the model prediction performance is related to the fact that the number of clients can also reduce the dataset size in each hospital.

4.5.2

Results for RQ2

Figure 8 shows the accuracy, precision, recall, and .F1 metrics with a different number of clients in the federated learning environment. As illustrated in the figure, the number of clients in the federated learning environment and prediction accuracy

Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in. . .

69

Fig. 7 The DL model’s prediction performance with a different number of IoT clients. (a) Video conferencing system. (b) Video conferencing system. (c) Video conferencing system. (d) Video conferencing system

Fig. 8 Prediction performance with the different number of clients

has a negative relationship. The final model’s classification performance appears to be decreasing as the number of clients grows. The figure shows that with increasing the number of clients, the DL model is becoming less accurate. However, with increasing the number of clients, the number of patients in the dataset is also decreasing.

70

K. Sahinbas and F. O. Catak

5 Threats to Validity A key external validity threat is related to the generalization of results. In our experiments, we used only one medical IoT dataset. We also definitely require more case studies to generalize the results. Besides, the dataset reflects ECG readings from human body. Our key construct validity threat is related to selecting the number of IoT devices that do not precisely show the DL model’s prediction performance. Nevertheless, note that the number of clients varies, and the DL models’ prediction performance is very close to each other. In the future studies, we will conduct more empirical studies to systematically investigate the effects of the number of clients in federated learning to the DL models’ prediction performance.

6 Conclusion Federated learning was mentioned as a new way to provide privacy preservation. The primary purpose of this chapter is to create a DL model for different hospitals based on the relevant patient records. Federated-learning-based privacy protection was proposed for medical IoT devices. This chapter contributes to explore privacypreserving of medical IoT devices and federated learning to protect privacy. We point out that the accuracy of the Bayesian NN model is roughly equal to the accuracy of the prediction performance of the base model. We indicate that the number of clients varies and the prediction performance of the DL models is very close to each other. As future research, we will implement partial and somewhat homomorphic encryption-based schemes to protect personal and sensitive data in IoT applications.

References 1. Aono, Y., Hayashi, T., Wang, L., Moriai, S., et al. (2017). Privacy-preserving deep learning via additively homomorphic encryption. IEEE Transactions on Information Forensics and Security, 13(5), 1333–1345. 2. Banos, O., Garcia, R., Holgado-Terriza, J. A., Damas, M., Pomares, H., Rojas, I., Saez, A., & Villalonga, C. (2014). mHealthDroid: A novel framework for agile development of mobile health applications. In Pecchia, L., Chen, L. L., Nugent, C., & Bravo, J., (Eds.), Ambient assisted living and daily activities (pp. 91–98). Cham: Springer International Publishing. 3. Bonawitz, K., Eichner, H., Grieskamp, W., Huba, D., Ingerman, A., Ivanov, V., Kiddon, C., Koneˇcný, J., Mazzocchi, S., McMahan, H. B., Van Overveldt, T., Petrou, D., Ramage, D., & Roselander, J. (2019). Towards Federated Learning at Scale: System Design. e-prints. arXiv:1902.01046. 4. Catak, F. O., Ahmed, J., Sahinbas, K., & Khand, Z. H. (2021). Data augmentation based malware detection using convolutional neural networks. PeerJ Computer Science, 7, e346.

Secure Multi-party Computation-Based Privacy-Preserving Data Analysis in. . .

71

5. Catak, F. O., Aydin, I., Elezaj, O., & Yildirim-Yayilgan, S. (2020). Practical implementation of privacy preserving clustering methods using a partially homomorphic encryption algorithm. Electronics, 9(2). 6. Chen, Y.-R., Rezapour, A., & Tzeng, W.-G. (2018). Privacy-preserving ridge regression on distributed data. Information Sciences, 451, 34–49. 7. Guo, X., Lin, H., Li, Z., & Peng, M. (2019). Deep-reinforcement-learning-based QoS-aware secure routing for SDN-IoT. IEEE Internet of Things Journal, 7(7), 6242–6251. 8. Haddadpour, F., & Mahdavi, M. (2019). On the convergence of local descent methods in federated learning. Preprint. arXiv:1910.14425. 9. Huang, C., Lee, H., & Lee, D. H. (2012). A privacy-strengthened scheme for e-healthcare monitoring system. Journal of Medical Systems, 36(5), 2959–2971. 10. Jiang, J., Ji, S., & Long, G. (2020). Decentralized knowledge acquisition for mobile Internet applications. World Wide Web (pp. 1–17). 11. Karimireddy, S. P., Kale, S., Mohri, M., Reddi, S., Stich, S., & Suresh, A. T. (2020). SCAFFOLD: Stochastic controlled averaging for federated learning. In International Conference on Machine Learning (pp. 5132–5143). PMLR. 12. Karimpour, N., Karaduman, B., Ural, A., Challenger, M., & Dagdeviren, O. (2019). IoT based hand hygiene compliance monitoring. In 2019 International Symposium on Networks, Computers and Communications (ISNCC) (pp. 1–6). IEEE. 13. Lalitha, A., Kilinc, O. C., Javidi, T., & Koushanfar, F. (2019). Peer-to-peer federated learning on graphs. Preprint. arXiv:1901.11173. 14. Li, T., Sahu, A. K., Zaheer, M., Sanjabi, M., Talwalkar, A., & Smith, V. (2018). Federated optimization in heterogeneous networks. Preprint. arXiv:1812.06127. 15. Lian, X., Zhang, C., Zhang, H., Hsieh, C.-J., Zhang, W., & Liu, J. (2017). Can decentralized algorithms outperform centralized algorithms? a case study for decentralized parallel stochastic gradient descent. Preprint. arXiv:1705.09056. 16. Lin, J., Yu, W., Zhang, N., Yang, X., Zhang, H., & Zhao, W. (2017). A survey on Internet of Things: Architecture, enabling technologies, security and privacy, and applications. IEEE Internet of Things Journal, 4(5), 1125–1142. 17. Lin, X., Lu, R., Shen, X., Nemoto, Y., and Kato, N. (2009). Sage: a strong privacy-preserving scheme against global eavesdropping for eHealth systems. IEEE Journal on Selected Areas in Communications, 27(4), 365–378. 18. Liu, L., Zhang, J., Song, S., & Letaief, K. B. (2020). Client-edge-cloud hierarchical federated learning. In ICC 2020-2020 IEEE International Conference on Communications (ICC) (pp. 1– 6). IEEE. 19. Liu, Y., Kang, Y., Zhang, X., Li, L., Cheng, Y., Chen, T., Hong, M., & Yang, Q. (2019). A communication efficient collaborative learning framework for distributed features. Preprint. arXiv:1912.11187. 20. Lu, R., Lin, X., & Shen, X. (2012). SPOC: A secure and privacy-preserving opportunistic computing framework for mobile-healthcare emergency. IEEE Transactions on Parallel and Distributed Systems, 24(3), 614–624. 21. Luo, S., Chen, X., Wu, Q., Zhou, Z., & Yu, S. (2020). HFEL: Joint edge association and resource allocation for cost-efficient hierarchical federated edge learning. IEEE Transactions on Wireless Communications, 19(10), 6535–6548. 22. Makhoul, J., Kubala, F., Schwartz, R., & Weischedel, R. (1999). Performance measures for information extraction. In Proceedings of DARPA Broadcast News Workshop (pp. 249–252). 23. Manning, C. D., Raghavan, P., & Schütze, H. (2008). Introduction to information retrieval. Cambridge, UK: Cambridge University Press. 24. McMahan, B., Moore, E., Ramage, D., Hampson, S., & y Arcas, B. A. (2017). Communicationefficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics (pp. 1273–1282). PMLR. 25. Nilsson, A., Smith, S., Ulm, G., Gustavsson, E., & Jirstrand, M. (2018). A performance evaluation of federated learning algorithms. In Proceedings of the Second Workshop on Distributed Infrastructures for Deep Learning (pp. 1–8).

72

K. Sahinbas and F. O. Catak

26. Pan, S. J., & Yang, Q. (2009). A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10), 1345–1359. 27. Sahinbas, K., & Catak, F. O. (2021). Transfer learning-based convolutional neural network for covid-19 detection with x-ray images. In Kose, U., Gupta, D., de Albuquerque, V. H. C., & Khanna, A., (Eds.), Data science for COVID-19 (pp. 451–466). Academic Press. 28. Turpin, A., & Scholer, F. (2006). User performance versus precision measures for simple search tasks. In Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’06, (pp. 11–18), New York, NY, USA: Association for Computing Machinery. 29. Vanhaesebrouck, P., Bellet, A., & Tommasi, M. (2016). Decentralized collaborative learning of personalized models over networks. Preprint. arXiv:1610.05202. 30. Vijayalakshmi, A. V., & Arockiam, L. (2018). A secured architecture for IoT healthcare system. In International Conference on Computer Networks, Big Data and IoT (pp. 904–911). Springer. 31. Wan, L., Ng, W. K., Han, S., & Lee, V. C. (2007). Privacy-preservation for gradient descent methods. In Proceedings of the 13th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 775–783). 32. Wang, H., Yurochkin, M., Sun, Y., Papailiopoulos, D., & Khazaeni, Y. (2020). Federated learning with matched averaging. Preprint. arXiv:2002.06440. 33. Wang, J., Sahu, A. K., Yang, Z., Joshi, G., & Kar, S. (2019). MATCHA: Speeding up decentralized SGD via matching decomposition sampling. In 2019 Sixth Indian Control Conference (ICC) (pp. 299–300). IEEE. 34. Yang, Q., Liu, Y., Chen, T., & Tong, Y. (2019). Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2), 1–19. 35. Yang, Y., Zheng, X., Guo, W., Liu, X., & Chang, V. (2018). Privacy-preserving fusion of IoT and big data for e-health. Future Generation Computer Systems, 86, 1437–1455. 36. Yu, F., Zhang, W., Qin, Z., Xu, Z., Wang, D., Liu, C., Tian, Z., & Chen, X. (2020). Heterogeneous federated learning. Preprint. arXiv:2008.06767. 37. Yu, L., Lu, Y., & Zhu, X. (2012). Smart hospital based on Internet of Things. Journal of Networks, 7(10), 1654. 38. Yuan, J., Xu, M., Ma, X., Zhou, A., Liu, X., & Wang, S. (2020). Hierarchical federated learning through LAN-WAN orchestration. Preprint. arXiv:2010.11612. 39. Zhang, A., Wang, L., Ye, X., & Lin, X. (2016). Light-weight and robust security-aware d2dassist data transmission protocol for mobile-health systems. IEEE Transactions on Information Forensics and Security, 12(3), 662–675.

A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos: A Covid-19 Solution Usman Ahmad Usmani, Junzo Watada, Jafreezal Jaafar, Izzatdin Abdul Aziz, and Arunava Roy

1 Introduction In December 2019, an infectious disorder linked to the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) was identified in Wuhan, Hubei, China [1, 2]. By August 23, 2020, over 23.3 million casualties had been recorded in 188 countries and territories, resulting in over 806,000 fatalities. 15 million people have been saved in particular [3]. Any national response to the Covid-19 pandemic has been driven by public opinion on social distance. Social distancing, also known as physical distancing, is a collection of methods used in Singapore to keep the virus from spreading from one individual to the next. Staying at home, driving from home, and preventing social activities are also main components of global public health advice. The urge to quit the house for essential jobs, food shopping, caring for sick and disabled people, and education is already evident at the national and international levels. People are told to maintain a certain minimum distance from others as they leave their homes, but some of the continuity of guidance stops here. Different nations have different minimum distances that people are encouraged to maintain. It ranges from 1 m, as recommended by WHO, to 1.5 m in Australia. The United States recommends 1.8 m (6 feet), while the United Kingdom, Ireland, and New Zealand both propose 2 m. To support pandemic preparedness, all governments argue that their advice is based

U. A. Usmani () · J. Jaafar · I. A. Aziz Universiti Teknologi Petronas (UTP), Seri Iskandar, Malaysia e-mail: [email protected]; [email protected] J. Watada Waseda University, Kitakyushu, Japan A. Roy Monash University Malaysia, Subang Jaya, Malaysia © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_4

73

74

U. A. Usmani et al.

Fig. 1 Social distancing graph

on reliable information. But how do different countries establish precise laws, and how should the general public know what the distances are? Modeling, simulation, and lessons learned from past virus outbreaks are used to develop guidelines for a safe spreading period. Recent work by Guo et al. indicates that Covid-19 is not entirely shielded for certain distances, with a range of up to 4 m found in a hospital setting, but it is uncertain if particles that travel so far are infectious [4]. The healthy distance recommendation is focused on the public’s understanding of the prescribed distance and capacity to accurately measure this distance from the patient. That will help the general public realize what they ought to do to stay accessible and secure in their daily lives (Fig. 1).

1.1 General Symptoms of Coronavirus Popular symptoms include fever, cough, nausea, shortness of breath, and a loss of smell and taste. Although the majority of people have just mild effects, others undergo acute respiratory distress syndrome (ARDS) and may be caused by cytokine storms, multi-organ failure, septic shock, or blood clots, among other things [5–8]. The time before the start of signs is normally about five days, although it may vary between two and fourteen days [9–11]. The most popular methods for the virus to spread through the nose and mouth are coughing, sneezing, and talking. Droplets should not be allowed to travel over lengthy periods of time in the normal course of events [12, 13]. However, those standing nearby can inhale these droplets

A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos. . .

75

[14] and become infected. People are more likely to get sick if they fall into contact with a polluted substance and rub their teeth. In confined environments, tiny droplets trapped in the breeze over lengthy stretches of time can scatter viruses [15–17]. It is most infectious in the first three days after symptoms appear, but it can also spread before symptoms start and among others who do not [18, 19]. The standard method of diagnosis is a real-time reverse transcription polymerase chain reaction (rRTPCR) from a nasopharyngeal swab [20, 21]. Although chest CT imaging may aid in the diagnosis of people at high risk of infection based on their symptoms and risk factors, it is not used for regular screening (Fig. 2). As the coronavirus travels across the world at an alarming rate, social exclusion seems to be the rule in many of the countries affected. While experts say it is too late to halt the pandemic, most policy policies are directed at “tracking the curve”

Fig. 2 Social distancing Statista (https://www.statista.com/chart/21129/google-searches-forsocial-distancing/)

76

U. A. Usmani et al.

or slowing the progression of the disease. As the graph below shows, several users are perplexed by the buzzword and are moving to Google for clarification. The global search volume for the word has risen to 100, according to Google Trends, with the two most popular search words being “coronavirus social distancing” and “social distancing meaning.” The Centers for Disease Control and Prevention (CDC) describes social distance as “the party has become isolated from its surroundings, that mass meetings must be avoided, and that it stays as far away from others as possible.”

1.2 Measures to Prevent Covid-19 Infection To avoid illness, hand washing, physical separation from others (especially those with symptoms), quarantine (especially those with symptoms), coughing, and holding unwashed hands away from the face are both recommended measures [22, 23]. Cotton face coverings, such as scarves or bandanas, were often proposed by health officials for use in public spaces to reduce the risk of pollution, with some authorities requesting their use. According to health officials, health face masks, such as N95 masks, are only permitted to be used by medical personnel, first responders, and those who are specifically caring for affected people [24, 25]. The World Health Organization (WHO) proclaimed the COVID19 outbreak a public health emergency of international significance (PHEIC) on 30 January 2020, and a pandemic on 11 March 2020 [26]. The disorder has spread locally in the majority of countries in all six WHO regions. COVID19 is a recent outbreak, and many aspects of its spread are still being studied [27]. It travels between people quickly, quicker than influenza but not as efficiently as measles. People are more contagious when they have symptoms (even though they are mild or non-specific), but they can remain contagious for up to two days until symptoms occur (pre-symptomatic transmission) [28]. In minor cases, they are infectious for seven to twelve days, while in extreme cases, they are contagious for two weeks on average. A June 2020 study found that 40–45% of the afflicted patients are asymptomatic, and the virus spreads mostly when individuals come into close contact and one human inhales tiny droplets produced by an infectious person (symptomatic or not) coughing, sneezing, smiling, or singing [29, 30]. The World Health Organization (WHO) suggests a social distance of one meter (3 feet), while the Centers for Disease Control and Prevention (CDC) in the United States recommends two meters [31]. According to the results of the tests, the virus will survive in aerosol for up to three hours. Some outbreaks have also been recorded in crowded and badly ventilated indoor areas where ill people spend long periods of time (such as restaurants and nightclubs). Of such cases, aerosol distribution has not been phased out. Smaller droplets will grow as a result of any surgical operations performed on COVID19 patients in hospitals [32] and result in the virus being transmitted more easily than normal.

A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos. . .

77

1.3 What Is Social Distancing? In public health, social distancing, also known as physical distancing, applies to a collection of non-pharmaceutical strategies or measures aimed at avoiding the spread of infectious diseases by maintaining a physical distance between persons and reducing the number of times people come into close contact with one another [33, 34]. It usually means having a certain space between them (the distance may be specified). By the likelihood of an uninfected person coming into close contact with an infected person, the disease’s spread may be stopped, resulting in less deaths. These can be used in conjunction with other procedures, such as adequate respiratory protection, facial masks, and hand washing. During the COVID-19 pandemic, the World Health Organization (WHO) proposed using the word “physical distance” rather than “social distance,” since physical separation prevents transmission; people may still be socially linked by technology. A variety of social distance interventions, such as school and office closures, segregation, quarantine, bans on citizen movement, and cancelation of mass gatherings, are used to slow the dissemination of contagious diseases and prevent overburdening healthcare systems, especially during a pandemic [35]. While the word “social-distancing programs” was not invented until the twenty-first century, they have existed since at least the fifth century BC. Community distancing methods have been employed extensively in a number of epidemics in recent years. Soon after the first influenza outbreaks were detected in St. Louis during the 1918 flu pandemic, officials enforced school closures, prohibitions on town meetings, and other social distance programs. St. Louis had significantly better influenza fatality rates than Philadelphia and has less influenza records that allowed for a public procession and did not institute social distancing for longer than two weeks after the first outbreaks [36]. Authorities facilitated or guided social exclusion during the COVID-19 pandemic. When infectious diseases transmit by one or all of the following means, such as droplet communication (coughing or sneezing), direct physical contact (including sexual intercourse), indirect physical contact (such as rubbing a contaminated surface), or airborne transmission (if the microorganism may stay in the air for an extended period of time), social distancing interventions are most efficient [37]. During the 2009 flu pandemic, the World Health Organization described social distancing as “holding at least an arm’s length away from others,[and] minimizing gatherings” [38]. When combined with good respiratory care and hand washing, it is considered to be the most practical way to prevent or postpone a pandemic [39]. During the COVID-19 pandemic, the Centers for Disease Control and Prevention redefined social distance as “staying out of congregate areas, avoiding mass meetings, and keeping distances (approximately six feet or two meters) from others wherever possible” [40]. The WHO uses the word “physical distance.” It is unclear why six foot was selected. According to new research, snow or heavy breathing droplets may be transferred over six meters during exercise [41]. According to researchers and scientists, greater social distinctions, as well as mask wear and social isolation, are needed.

78

U. A. Usmani et al.

Our main contributions include: • A deep-learning-based social distancing tool that can be used to monitor social distance being followed up in real-time videos. • This method works well in congested cities, encouraging authorities to take appropriate action against those that violate social distancing norms. • There is no other social distancing tool that can monitor the follow-up, and this tool works well in many complex situations. Holding a constant physical distance from one another and avoiding embraces and movements that require direct physical contact decrease the risk of infection during infectious respiratory disease outbreaks (e.g., flu pandemics and the 2020 pandemic of COVID-19). Separation lengths and personal grooming precautions are often recommended in the workplace. During the COVID-19 pandemic, for example, the World Health Organization recommends a distance of at least 1 m. (3.3 ft). Following that, Australia, Denmark, Hong Kong, Lithuania, and Singapore all enacted a one-million-dollar social distancing plan. In South Korea, a 1.4square-kilometer area has been established (4 ft 7 in). Australia, Belgium, Germany, Greece, Italy, the Netherlands, Portugal, and Spain are among the countries that have taken 1.5 million inhabitants (4.9 ft). The United States has set an overall height limit of 6 feet (1.8 m), whereas Canada has set a limit of 2 meters (6.6 ft). While the UK initially recommended a gap of at least 2 m, as of July 4, 2020, this was reduced to a "1 m+" solution, reducing the safe distance to 1 m (3.3 ft) where other virus-prevention measures, such as face masks or plastic windows, were in use. The WHO’s one-meter recommendation is for tuberculosis transmitting testing involving William F. Wells droplets. An analysis of SARS transmission through aircraft published in The New England Journal of Medicine may have influenced the 6 ft (1.8 m) CDC. Despite this, the CDC offered no valuable information when approached [40]. Any drills for handshakes have been suggested. One non-touch choice is the namaste sign, which involves placing one’s palms together, pointing fingers upward, and pushing one’s hands to the center. According to WHO Director-General Dr. Tedros Adhanom Ghebreyesus and Israeli Prime Minister Benjamin Netanyahu, Prince Charles made this gesture at a guest reception during the COVID-19 pandemic in the United Kingdom [42]. The thumbs up, shake, shaka (or “keep loose”) sign, and palm on the forehead are all popular movements. Previous influenza pandemics, such as the 2009–2010 pandemic, have shown that the next pandemic would not be able to be tracked internationally, nor will we be able to deter foreign virus spread for more than a short time. Antiviral drug availability could be limited in the early stages of the next pandemic, with the bulk of antiviral drugs intended for more acute diseases and patients at higher risk of influenza.

A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos. . .

79

1.4 Social Distancing as a Tool to Prevent the Virus As a result, health officials will focus heavily on non-pharmaceutical approaches (NPIs), such as social distancing, in order to limit influenza spread in the community and achieve three particular outcomes. The first result would be a postponement of the peak of the outbreak to allow for surgical preparations, the second would be a reduction in the peak of the epidemic to avoid overburdening the healthcare system, and the third would be a scattering of outbreaks over longer periods of time, allowing for easier management of those situations and the use of vaccines at least later in the epidemic. Influenza infections are expected to spread mostly by close contact throughout the population (e.g., households, businesses, nursing facilities, hospitals, and public places), with children playing a special role in transmission. Interventions that encourage social distancing seek to minimize touch frequency and increase physical distance between individuals, thus lowering the probability of personal transmission. Similar initiatives assisted in the relief of earlier pandemics, such as the 1918–1919 pandemic, which continue to play an important role in modern pandemic preparedness strategies. Despite the fact that prospective social isolation interventions have a solid biological and epidemiological foundation, there are few resources for systematic randomized trials of community influenza approaches. The aim of this chapter was to examine the evidence for social distancing techniques, with a particular emphasis on evidence that they are effective at minimizing group influenza transmission.

2 Related Work Our specific approach is based on the idea of Sensitive Infected Recovered (SIR). A limited amount of research were conducted prior to COVID-19 with the aim of incorporating optimum economic or social behavior regulation into the disease dynamic model [43]. Prior to 2020, comparative epidemiological models with traditional economic features will be conducted. Despite the fact that we looked at the evidence for increasing NPI separately, it is normal to combine social distancing interventions. Several NPIs were enforced concurrently throughout a variety of cities in the United States during the 1918 influenza pandemic, including school closings and bans on public collecting. Though simulation studies projected that if more NPIs were introduced, we agree that some consideration should be provided to finding approaches that complement each other when used together. The introduction of social distance-creating practices such as school and mall openings at the same time may avoid a rise in social contact outside of schools. School closures can be combined with phone-based options that enable parents to care for their children when they are at school. Given the limits and contradictions, attempts to separate social security from the next pandemic would be critical components of the public policy approach. These

80

U. A. Usmani et al.

practices must be thoroughly examined when implementing pandemic strategies, particularly in terms of law enforcement and resource planning and distribution. The suggestion that sick people remain at home is perhaps the most successful social distance mechanism, and pandemic preparation should consider if sick children and employees will stay at home from school or work. Health departments, for example, may waive the standard requirement for medical notes to support school or work absences. Finally, although our research focused on non-pharmacological measures to take during influenza pandemics, the results may also be extended to extreme seasonal influenza outbreaks. Finally, our research supports the effectiveness of social distance intervention during influenza pandemics based on observer and simulation results. These community-wide programs will benefit from improved implementation and compliance if they are implemented on time. More dissemination experiments, as well as studies into the appropriate times and lengths for school and workplace closures, will be helpful. The only feasible and timely way to produce such evidence is to simulate the mathematical replication of a novel pathogen, such as SRAS CoV-2, under various scenarios. Koo and colleagues3 used granular data on the composition and behavior of the Singaporean population to measure the possible effect of special social-distancing measures on SARS-CoV-2 complex transmission. The researchers identified three degrees of infectivity (15, 20, or 25 simple numbers [R0]) and predicted that 75 to 50% of the citizens will be asymptomatic. With or without school openings and office separations, quarantine was used for the remainder of the steps (50% of the telecommuting workers). Despite the model’s configuration making interpreting the impact of each parameter difficult, the key results for sensitivity analyses remain robust. When R0 was 15, the combined action reduced the estimated median amount of pathogens by 993% (IQR 926–999); when R0 was 20, the combined action reduced the estimated median amount of pathogens by 782% (590–944); and when R0 was 25, the combined action reduced the estimated median amount of pathogens by 782% (590–944). It is unsurprising that the combined action culminated in the largest drop in COVID-19 incidents. The measurement of each operation’s added importance, on the other side, offers critical details. Since each response has the potential to cause substantial social harm, determining how much work is needed to minimize disease transmission and burden is important. On a daily basis, new hypotheses about the SARS-CoV-2 transmission mechanism and clinical profile arise, including the likelihood of a significantly underestimated child infection rate8. Estimates for Singapore’s reproductive numbers, however, are not yet accessible. Although asymptomatic infections are uncommon, the authors concluded that 75% of the infections are clinically asymptomatic; this importance will affect the efficacy of social distance therapies, as shown by Koo and colleagues in sensitivity analyses of higher asymptomatic proportions. Furthermore, it implies that the general public has a high degree of loyalty, which cannot be guaranteed. While the scientific rationale for these approaches is convincing, there are several ethical concerns. 9 Government authorities must implement non-discriminatory quarantine and welfare-distant policies for all racial groups. The repercussions of social and economic inequality perpetrated in the name of public health are far-reaching. 10

A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos. . .

81

interventions may result in reduced wages and even employment displacement, affecting the most vulnerable populations more. Particular consideration should be paid to the protection of marginalized persons such as the homeless, the incarcerated, the aged or the disabled, and illegal immigrants. The feasibility and cultural effects of quarantine and social distancing can be determined by the authority of public health officials, political leaders, and organizations. Public health agencies, administrators, and organizations will decide the effects of quarantine and social isolation. To maintain public trust, policymakers must use evidence-based approaches and fully free, factual debate. An epidemiological model and mathematical modeling approaches from economics were used to comment on different aspects of COVID-19. According to these articles, we are focusing on building the strongest possible model that can quantitatively and empirically capture the costs and benefits of widespread social distance. The quantitative goal keeps us centered on the fact that the disease can be spread across a wide range of social interactions, only a subset of which can be observed through intake or labor supply. Fortunately, we can now map our model to recently available high-frequency social interaction data from Google and SafeGraph. This form of data was only made available to researchers after the start of COVID-19. Around the same time, our focus on social interaction enables us to link our model to current policy, social distancing. Our methodology is perhaps most unique in that we have a collection of data on different aspects of both human behavior and disease characteristics that we use to quantify the model’s results. The evidence is analyzed not only in terms of social distance, but also of mortality and the effective reproduction number R(t). As a result, we use the model’s potential to clarify why laissez-faire and optimal methods produce the results that they do. Since the models are so identical, our paper draws qualitative results that are very similar to those of other articles in the growing COVID-19 literature. Individual motivation is helpful in reducing societal alienation, but it is not enough. The best strategy is to erect immediate social barriers, which will enable the virus to spread before herd immunity is achieved. In these papers, we evaluate the model’s predictive efficiency using a large amount of data on different facets of the epidemic. In these articles, we provide a consistent and practical guide to the study of both balance and equilibrium, provide new proof of balance activity in the early stages of the COVID-19 outbreak, and provide a quantitative assessment of the model’s predictive performance based on a wealth of data on different aspects of the epidemic. Our quantitative approximation that an optimal solution is consistent but not too rigid is special in the literature. We claim to be the first to demonstrate a connection between these optimal strategy characteristics and the effective reproduction number R(t), a crucial epidemiological phenomenon, which stabilizes about 1. Our paper’s information arrangement varies from that of the others on a larger scale. Recently infected people are believed to be ignorant of their infection status and, as a consequence, inadvertently transmit the disease. Due to the failure of the procedure to distinguish between susceptible and contagious individuals, social distance techniques must be large rather than targeted at infected people. Typically, these papers include a lot of observations

82

U. A. Usmani et al.

regarding people’s knowledge bases. Any posts on the same subject include a wide range of policy choices. They wanted to concentrate on pervasive social isolation because it was prevalent early in the COVID-19 outbreak and is likely to recur in subsequent outbreaks. In practice, advanced technology such as testtrace-and-quarantine is often inaccessible or ineffective months after an outbreak. Diamond and Maskin [44] wrote regarding social externalities and differentiated the quadratic and linear matching technologies. In quadratic mixing, increased physical interaction with others increases the probability of social contact and, as a result, disease transmission among all participants. Since parks, malls, and mass transportation are all more busy these days, each trip to the park/restaurant/subway is more likely to result in illness. A corresponding function with such a search externality has a search externality that is generally regarded as positive in the sense of the disorder. In contrast, in a linear search system, an individual’s overall social interaction is primarily dictated by his or her own social context, rather than by the actions of others. This technology, we think, can be utilized in contexts where individuals are actively pursuing social relationships. As a consequence, we conclude that quadratic technology is adequate for modeling COVID-19 dynamics, while linear technology could be optimal for modeling an infection such as HIV. Social distancing at work, according to epidemiological and statistical findings, reduces the cumulative number of influenza events. The influenza outbreak’s staggered peak brought it all to a head. Influenza prevention is more successful when paired with other occupational therapy, according to the modeling findings. Higher R0 values, a delayed social inequality, or lower enforcement, on the other hand, are expected to limit efficacy. Droplets and perhaps aerosols are present in coughs and sneezes, which are the most popular forms for the flu to spread [45]. Social isolation at work decreases the risk of person-to-person transmission of influenza by limiting gout transmission within 3–6 ft [46]. Social isolation from the workplace, along with other non-pharmaceutical or pharmaceutical therapies, can aid in reducing virus spread [47]. At higher R0 values, social distancing was found to be less successful. If R0 is greater, efficiency will be smaller, and social distancing will reduce the potency of reproduction to less than one. The poorer outcomes of delayed enforcement may be due to a number of factors. Distancing oneself socially from work restricts one’s ability to impact existing cases and provides a missed opportunity to restrict further distribution. Specific transmission becomes more likely as enforcement grows. This systematic analysis could have certain drawbacks. To continue, the majority of the experiments used were modeled, and none were actively conducted. Models will fill holes in decision-making where experience is lacking. However, more epidemiological evidence on social distancing is required in specific contexts. Furthermore, the effectiveness of the simulation experiments has not been evaluated. Population characteristics identifying exposure points (e.g., in the family, school, or workplace), population characteristics representing degrees of exposure (e.g., contact rates and duration), and disease-transmitting parameters are among the model simulation input parameters. Several empirical studies on the rates of workplace contact have been conducted. There has been no long-term study of the impact of social distance methods in the workplace on increasing contact

A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos. . .

83

rates. Third, the included study did not consider the impact of occupational social exclusion on two of our key outcomes of importance (lost working days, damages). The effect on missed business days will be a trade-off between possible labor loss due to social distance (which can be mitigated by the opportunity to work from home) and sick days saved due to decreased influenza spread and sickness. According to the report, Hispanic and African American workers would outnumber white workers.

3 Methodology At this stage, social isolation is the best way to prevent the spread of COVID-19. Detectron 2 is an open-source library for object detection and segmentation created by Facebook’s FAIR AI development team. For the resolution of different computer vision functions, Detectron 2 incorporates cutting-edge architectures such as Faster RCNN, MaskR CNN, and RetinaNet. To produce these so-called proposals, a small network is slid over a convolutional feature map generated by the final convolutional layer for the area in which the object is located. The proposal for the artifacts is generated by RPN. RPN has a one-of-a-kind and cutting-edge architecture. RPN includes both a classifier and a regressor. The writers invented the term “anchors.” The anchor is the focal point of the sliding glass. The ZF Model, an AlexNet expansion, has 256-d dimensions, while the VGG-16 has 512-d dimensions. The classifier determines the likelihood of a proposal possessing the target item. The location of the hypotheses is regressed using regression. The scale and aspect ratio of a photograph are two critical considerations. For those that do not know, aspect ratio = width of picture/height of image, and scale is the image’s size. These anchors are classified according to two criteria: Anchors with the largest intersection-over-union value congregate on a ground reality box. Anchors with a greater-than-0.7 intersection-over-union overlap. RPN is, finally, a procedural algorithm that must be perfected. As a consequence, what we have is a loss function. https://medium.com/egen/region-proposal-network-rpn-backbone-of-faster-rcnn-4a744a38d7f9 Detectron is a software framework developed by Facebook AI Research (FAIR) that implements cutting-edge object detection algorithms such as Mask R-CNN. It is written in Python and comes with the Caffe2 deep learning platform. Pyramid feature networks, R-CNN mask, human object detection and recognition, dense object detection focal loss, non-locomotive neural networks, segment training, and information distillation, toward omni-supervised training, were among the research projects proposed by Detectron at FAIR. Mask RCNN is used for object identification.

84

U. A. Usmani et al.

The RCNN is a deep neural network programmed to solve instance segmentation problems in machine learning or computer vision. To put it another way, various elements in a picture or video may be distinguished. When you are given a picture, it provides you with object bounding boxes, classes, and masks. The RCNN mask is split into two levels. It first produces recommendations for regions in which an object should be located based on the input image. Second, it predicts the object type, refines the boundary box, and produces a pixel-level mask dependent on the proposal for the first step. The backbone is a deep neural FPN network. Both points are connected by the backbone mechanism. It is divided into three sections: bottom-up, top-down, and side ties. Any ConvNet that extracts raw image attributes, typically ResNet or VGG, may be used as the bottom-up path. The top–bottom path creates a pyramid map of the same size as the bottom-up path. Lateral relations are transfers and additional activities that take place in two directions. Because of its ability to retain good semantic properties at different resolution scales, FPN outperforms other individual ConvNets. An RPN, which is a lightweight neural network, scans all top–bottom FPN pathways and suggests areas that might contain artifacts. That is what there is to it. While scanning the map is a quick operation, we will need a way to associate characteristics with the raw shot. The anchors also arrived. Anchors are a group of predefined boxes with predefined image sizes and positions. Individual anchors with a particular IoU attribute are allocated ground truth groups and bounding boxes (at this point, only binary items or contexts are specified). Since anchors with different scales are related to different feature map sizes, RPN can use anchors with different scales to decide where an entity is on the character map and the size of its bounding box. We both accept that convolving, sampling, and up-sampling will keep functions in the same relative locations as the points in the original image and have no impact on them. In the second level, a different neural network assigns the proposed regions to several distinct areas on the characteristics map, scans specific areas, and produces item groups (multicategories). The treatment seems to be RPN-based. Step two, on the other hand, uses a methodology known as ROIAlign to label the respective attribute map areas without the use of anchors, and each object has pixel-level branch generation masking. Mask-RCNN is a 2017 model that expands the faster-RCNN model for semantical segmentation, entity localization, and natural picture object instance segmentation. The authors define Mask-RCNN as a “simple, scalable, and general framework for entity instance segmentation.” Mask-RCNN was used to beat all existing single model entrants in the 2016 COCO Competition, overcoming a major barrier in object identification, segmentation, and underwriting. Many existing image segmentation algorithms may be divided into two types: those that use an area proposal algorithm and those that do not. U-Net, for example, is a segmentation algorithm that does not use a region proposal algorithm; instead, U-Net uses an encoder–decoder scheme in which a convergence neural network recognizes, or codes, the performance of the picture representation, and a second network, such as the disconvolutionary neural network, produces the desired segmentation mask from the learned representativeness. Encoder–decoder architectures have long been used in machine learning for tasks other than goal identification and segmentation,

A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos. . .

85

such as image denoising and generation. In contrast, Mask-RCNN is concerned with regional proposals produced by a network of regional proposals. Mask-RCNN is built on the faster-RCNN extractor model, and this area proposal network is followed by an operation called ROI-Pooling, which has three main adjustments to generate uniform size outputs for classification input. Mask-RCNN first replaces an operation named ROIAlign with one that allows for the creation of precise instance segmentation masks in an imprecise manner, and then to achieve the desired instance segmentation, mask-RCNN adds a network head. The mask and class predictions are eventually unbundled, and the mask network head forecasts the mask independently of the forecasting network head. L = Lcls + Lbbox + Lmask is a failure function that performs several tasks. Mask-RCNN is a backbone neural framework that is used to erase features. The backbone network may be any convolutional neural network built to process images, such as ResNet-50 or ResNet-101, but it has been shown that a function pyramid network (FPN) based on a network such as ResNet-50 or ResNet-101 offers precision and speed gains to the mask-RCNN. A function pyramid network employs the intrinsic hierarchical and multi-scale form of convolutional neural networks to generate useful features at various scales for object identification, semantic segmentation, and instance segmentation. Both function pyramid network models include a “backbone” network for the construction of the task pyramid. Again, the backbone model is typically a convolutional neural network, which is known for its high recognition efficiency and ability to be pretrained. The usefulness of Mask RCNN should be taken into account in image segmentation, as well as the fact that the properties of natural photos are usually substantially different from those of medical pictures [48]. If we know the bounding box of each person, we can easily determine their distance from us. The challenge here is determining the best coordinate to represent a human as a rectangular box. Make a function that returns the center of each boundary box. Calculate the center of each boundary box’s base and label it with a dot on the diagram. I choose the bottom center of the rectangle to correctly represent each entity, and this measurement is also invariant of a person’s height. Defines a function that computes the Euclidean distance between two points in a picture and returns the people who are closest to the given distance. In this scenario, the proximity distance is the shortest distance between two individuals. I have decided to be here till I am 100. Remember the people in the immediate vicinity. The findings show that four people reach the red zone when the difference is less than the proximity threshold. Defines a method for shifting the color of the people closest to you to red. Brighten the color of the ones closest to you. So far, we have seen a step-by-step method for calculating the distance between each pair of individuals using Detectron-2’s goal detection. These steps must now be repeated with each frame of the video. Defines a function that executes all steps on each frame of the film. Adjust the color of the video to red after identifying the closest figures in each picture.

86

Fig. 3 Mask-RCNN architecture (taken from medium.com)

Fig. 4 Multiple objects being detected by the model

U. A. Usmani et al.

A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos. . .

87

Fig. 5 Drawing a bounding box for one of the people

Fig. 6 Red boxes indicate the social distancing not being followed. Blue boxes indicate the followup of social distancing

4 Conclusions It is difficult to incorporate therapeutic distancing techniques. These would more likely be used during the local epidemic and will most likely be used before the manufacturing and distribution of a strain-specific vaccine is done. If the response is good enough throughout this time frame, a neighborhood outbreak can be

88

U. A. Usmani et al.

avoided. Nonetheless, affected regions may continue to incorporate flu and spread the local influenza, but they are more easily accommodated by health systems at a suffering level if neighboring populations do not even follow these interventions. The association between the disease host and the social networking network in which the disease spreads is central to our design strategy. Measuring community communication networks for infectious disease transmission necessitates concentrated study that incorporates sociology, public health, and epidemiology. These networks would most certainly vary based on whether the population is urban, rural, or limited. A network for any population of concern may be built with the aid of robust longitudinal evidence, trained elicitation of social scientists and community representatives, behavioral experiments, and prospective trials. Since young people had the highest death rate of any demographic during the 1918–19 Spanish pandemic, configurations that take into consideration, for example, school campuses or military reserves, might be beneficial.

References 1. Hui, D. S., Azhar, E. I., Madani, T. A., Ntoumi, F., Kock, R., Dar, O., Ippolito, G., Mchugh, T. D., Memish, Z. A., Drosten, C., Zumla, A., & Petersen, E. (2020). The continuing 2019nCoV epidemic threat of novel coronaviruses to global health — The latest 2019 novel coronavirus outbreak in Wuhan, China. International Journal of Infectious Diseases, 91, 264– 266. https://doi.org/10.1016/j.ijid.2020.01.009. 2. Lu, H., Stratton, C. W., & Tang, Y. W. (2020). Outbreak of pneumonia of unknown etiology in Wuhan, China: The mystery and the miracle. Journal of Medical Virology, 92(4), 401–402. https://doi.org/10.1002/jmv.25678. 3. Huang, C., Wang, Y., Li, X., Ren, L., Zhao, J., Hu, Y., Zhang, L., Fan, G., Xu, J., Gu, X., & Cheng, Z. (2020). Clinical features of patients infected with 2019 novel coronavirus in Wuhan. China. The Lancet, 395(10223), 497–506. 4. Mehta, P., McAuley, D. F., Brown, M., Sanchez, E., Tattersall, R. S., & Manson, J. J. (2020). COVID-19: consider cytokine storm syndromes and immunosuppression. The Lancet, 395(10229), 1033–1034. https://doi.org/10.1016/s0140-6736(20)30628-0. 5. WHO. (2020). Advice on the use of masks for children in the community in the context of COVID-19: annex to the advice on the use of masks in the context of COVID-19. 6. Ogundokun, R. O., Lukman, A. F., Kibria, G. B., Awotunde, J. B., & Aladeitan, B. B. (2020). Predictive modelling of COVID-19 confirmed cases in Nigeria. 7. Park, M., Won, J., Choi, B. Y., & Lee, C. J. (2020). Optimization of primer sets and detection protocols for SARS-CoV-2 of coronavirus disease 2019 (COVID-19) using PCR and real-time PCR. Experimental & Molecular Medicine, 52(6), 963–977. https://doi.org/10.1038/s12276020-0452-7 8. Lai, C. C., Liu, Y. H., Wang, C. Y., Wang, Y. H., Hsueh, S. C., Yen, M. Y., Ko, W. C., & Hsueh, P. R. (2020). Asymptomatic carrier state, acute respiratory disease, and pneumonia due to severe acute respiratory syndrome coronavirus 2 (SARSCoV-2): facts and myths. Immunology and Infection, 53(3), 404–412. 9. Furukawa, N. W., Brooks, J. T., & Sobel, J. (2020). Evidence supporting transmission of severe acute respiratory syndrome coronavirus 2 while presymptomatic or asymptomatic. Emerging Infectious Diseases, 26(7), 26–26. https://doi.org/10.3201/eid2607.201595. 10. Chau, N. V. V., Dung, N., Yen, L. M., Minh, N., & Hung, L. M. (2020). The natural history and transmission potential of asymptomatic SARS-CoV-2 infection. Clinical Infectious Diseases. https://doi.org/10.1101/2020.04.27.20082347. 11. Gandhi, R. T., Lynch, J. B., & del Rio, C. (2020). Mild or moderate Covid-19. New England Journal of Medicine. https://doi.org/10.1056/nejmcp2009249.

A Deep Learning Algorithm to Monitor Social Distancing in Real-Time Videos. . .

89

12. Oxley, T. J., Mocco, J., Majidi, S., Kellner, C. P., Shoirah, H., Singh, I. P., Leacy, R. A. D., Shigematsu, T., Ladner, T. R., Yaeger, K. A., Skliut, M., Weinberger, J., Dangayach, N. S., Bederson, J. B., Tuhrim, S., & Fifi, J. T. (2020). Large-vessel stroke as a presenting feature of Covid-19 in the young. New England Journal of Medicine, 382(20), e60–e60. https://doi.org/ 10.1056/nejmc2009787. 13. Rothan, H. A., & Byrareddy, S. N. (2020). The epidemiology and pathogenesis of coronavirus disease (COVID-19) outbreak. Journal of Autoimmunity, 109, 102433–102433. https://doi.org/ 10.1016/j.jaut.2020.102433. 14. Patel, R., Babady, E., Theel, E. S., Storch, G. A., Pinsky, B. A., George, K. S., Smith, T. C., & Bertuzzi, S. (2020). Report from the American Society for Microbiology COVID19 international summit. 15. Recalcati, S. (2020). Cutaneous manifestations in COVID-19: a first perspective. Journal of the European Academy of Dermatology and Venereology, 34(5). https://doi.org/10.1111/jdv. 16387. 16. Stadnytskyi, V., Bax, C. E., Bax, A. & Anfinrud, P. (2020). The airborne lifetime of small speech droplets and their potential importance in SARS-CoV-2 transmission. Proceedings of the National Academy of Sciences, 117(22), 11875–11877. 17. Wang, J., & Du, G. (1971). COVID-19 may transmit through aerosol. Irish Journal of Medical Science, 1–2. 18. Bahl, P., Doolan, C., de Silva, C., Chughtai, A. A., Bourouiba, L., & MacIntyre, C. R. (2020). Airborne or droplet precautions for health workers treating Coronavirus disease 2019? The Journal of Infectious Diseases. https://doi.org/10.1093/infdis/jiaa189. 19. Singhal, T. (2020). A review of coronavirus disease-2019 (COVID-19). The Indian Journal of Pediatrics, 1–6. 20. Roser, M., Ritchie, H., Ortiz-Ospina, E., & Hasell, J. (2020). Coronavirus pandemic (COVID19). Our World in Data. 21. Gefen, A., & Ousey, K. (2020). Update to device-related pressure ulcers: SECURE prevention. COVID-19, face masks and skin damage. Journal of Wound Care, 29(5), 245–259. https://doi. org/10.12968/jowc.2020.29.5.245. 22. Qaseem, A., Etxeandia-Ikobaltzeta, I., Yost, J., Miller, M.C., Abraham, G. M., Obley, A. J., Forciea, M. A., Jokela, J. A., & Humphrey, L. L. (2020). Use of N95, surgical, and cloth masks to prevent COVID-19 in health care and community settings: living practice points from the American College of Physicians (version 1). 23. Feng, S., Shen, C., Xia, N., Song, W., Fan, M., & Cowling, B. J. (2020). Rational use of face masks in the COVID-19 pandemic. The Lancet Respiratory Medicine, 8(5), 434–436. https:// doi.org/10.1016/s2213-2600(20)30134-x. 24. Lawand, N. M., & Tabbah, A. S. (2020). Coronavirus disease 2019 (COVID-19): prevention and disinfection. 25. Bauwens, M., Compernolle, S., Stavrakou, T., Müller, J. F., van Gent, J., Eskes, H., Levelt, P. F., Van Der A, R., Veefkind, J. P., Vlietinck, J., Yu, H., & Zehner, C. (2020). Impact of Coronavirus outbreak on No 2 pollution assessed using TROPOMI and OMI observations. Geophysical Research Letters, 47(11), e2020GL087978. https://doi.org/10.1029/2020gl087978. 26. Lax, S. F., Skok, K., Zechner, P., Kessler, H. H., Kaufmann, N., Koelblinger, C., Vander, K., Bargfrieder, U., & Trauner, M. (2020). Pulmonary arterial thrombosis in COVID-19 with fatal outcome: results from a prospective, single-center, clinicopathologic case series. Annals of Internal Medicine. https://doi.org/10.7326/m20-2566. 27. Tindale, L., Coombe, M., Stockdale, J. E., Garlock, E., Lau, W. Y. V., Saraswat, M., Lee, Y. H. B., Zhang, L., Chen, D., Wallinga, J., & Colijn, C. (2020). Transmission interval estimates suggest pre-symptomatic spread of COVID-19. 28. Kong, D., Zheng, Y., Wu, H., Pan, H., Wagner, A. L., Zheng, Y., Gong, X., Zhu, Y., Jin, B., Xiao, W., & Mao, S. (2020). Pre-symptomatic transmission of novel coronavirus in community settings. Influenza and Other Respiratory Viruses, 14(6), 610–614. 29. Somsen, G. A., van Rijn, C., Kooij, S., Bem, R. A. & Bonn, D. (2020). Small droplet aerosols in poorly ventilated spaces and SARS-CoV-2 transmission. The Lancet Respiratory Medicine, 8(7), 658–659.

90

U. A. Usmani et al.

30. Asadi, S., Wexler, A. S., Cappa, C. D., Barreda, S., Bouvier, N. M., & Ristenpart, W. D. (2019). Aerosol emission and superemission during human speech increase with voice loudness. Scientific Reports, 9(1), 1–10. https://doi.org/10.1038/s41598-019-38808-z. 31. Rajendra Prasad, N. L. (2020). Causes and consequences of COVID-19 on Society-A sociological study. IJRAR-International Journal of Research and Analytical Reviews, 2348– 1269. 32. Lai, M. Y. Y., Cheng, P. K. C., & Lim, W. W. L. (2005). Survival of severe acute respiratory syndrome coronavirus. Clinical Infectious Diseases, 41(7), e67–e71. https://doi.org/10.1086/ 433186. 33. European Centre for Disease Prevention and Control. (2008). How to protect yourself & others. Centers for Disease Control and Prevention. 34. Glass, R., Glass, L., Beyeler, W., & Min, H. (2006). Targeted social distancing designs for pandemic influenza. Emerging Infectious Diseases, 12(11), 1671–1681. https://doi.org/10. 3201/eid1211.060255. 35. Greenstone, M., & Nigam, V. (2020). Does social distancing matter? Becker Friedman Institute for Economics Working Paper, 2020–2046. 36. Fong, M. W., Gao, H., Wong, J. Y., Xiao, J., Shiu, E. Y., Ryu, S., & Cowling, B. J. (2020). Nonpharmaceutical measures for pandemic influenza in nonhealthcare settings—social distancing measures. Emerging Infectious Diseases, 26(5), 976–984. https://doi.org/10.3201/ eid2605.190995. 37. Lewnard, J. A., & Lo, N. C. (2020). Scientific and ethical basis for social-distancing interventions against COVID-19. https://doi.org/10.1016/s1473-3099(20)30190-0. 38. Wilder-Smith, A., & Freedman, D. O. (2020). Isolation, quarantine, social distancing and community containment: pivotal role for old-style public health measures in the novel coronavirus (2019-nCoV) outbreak. Journal of Travel Medicine, 27(2), 20–20. https://doi.org/ 10.1093/jtm/taaa020. 39. Martin-Loeches, I., Torres, A., Rinaudo, M., Terraneo, S., de Rosa, F., Ramirez, P., Diaz, E., Fernández-Barat, L., Bassi, G. L. L., & Ferrer, M. (2015). Resistance patterns and outcomes in intensive care unit (ICU)-acquired pneumonia. Validation of European Centre for Disease Prevention and Control (ECDC) and the Centers for Disease Control and Prevention (CDC) classification of multidrug resistant organisms. Journal of Infection, 70(3), 213–222. https:// doi.org/10.1016/j.jinf.2014.10.004. 40. Kim, S., & Liu, B. F. (2012). Are all crises opportunities? A comparison of how corporate and government organizations responded to the 2009 flu pandemic. Journal of Public Relations Research, 24(1), 69–85. 41. Kaushal, D., Nair, N. P., Soni, K., Goyal, A., Choudhury, B., & Rajan, N. (2020). Endoscopy in otorhinolaryngology during Corona outbreak: a proposal for safe practice. Indian Journal of Otolaryngology and Head & Neck Surgery, 1–6. https://doi.org/10.1007/s12070-020-02048-9. 42. Hopkins, C., Surda, P., & Kumar, N. (2020). Presentation of new onset anosmia during the COVID-19 pandemic. 43. Philipson, T. (2000). Economic epidemiology and infectious diseases. Handbook of health economics, 1, 1761–1799. 44. Diamond, P. A., & Maskin, E. (1979). An equilibrium analysis of search and breach of contract, I: Steady states. The Bell Journal of Economics, 282–316. 45. Yang, S., Lee, G. W., Chen, C. M., Wu, C. C., & Yu, K. P. (2007). The size and concentration of droplets generated by coughing in human subjects. Journal of Aerosol Medicine, 20(4), 484–494. 46. Thiele, R. G. (2011). Role of ultrasound and other advanced imaging in the diagnosis and management of gout. Current Rheumatology Reports, 13(2), 146–153. https://doi.org/10.1007/ s11926-010-0156-4. 47. Cavallaro, J. J., & Monto, A. S. (1970). Community-wide outbreak of infection with a 229Elike coronavirus in Tecumseh, Michigan. https://doi.org/10.1093/infdis/122.4.272. 48. He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask R-CNN. In Proceedings of the IEEE International Conference on Computer Vision (pp. 2961–2969).

Osteoarthritis Detection and Classification in Knee X-Ray Images Using Particle Swarm Optimization with Deep Neural Network N. Hema Rajini and A. Anton Smith

1 Introduction The Internet of Healthcare Things (IoHT) is a concept that defines exclusively identifiable devices linked to the Internet and has the capability of communicating with one another employed in the healthcare sector [1, 2]. Remote or automated management of resources can result in improved healthcare services and cost effectiveness. The IoHT manages the healthcare services and offers continual access to the equipment, data, and patient information. At the same time, explainable artificial intelligence (XAI) techniques can be designed to assist physicians in the healthcare sector [3, 4]. Osteoarthritis (OA) is an unbearable and expensive disease to every person present in the world [5]. Since a person lives longer and the world’s population is drastically increased, the medication cost is rising, and OA has become a significant problem that could not be eliminated. A better method to mitigate this cost is by the use of earlier identification, and, prominently, forecast of OA. To attain it, new automatic identification and diagnosis models are developed for the assistance of physicians. These models will enable the detection of patients on the growth way to OA. In the case of individual people, managing OA can undergo initialization in a timely way, hence, minimizing the influence of OA on the quality of living and reducing growth. The patients can also get an advantage by including the medical examination on disease-changing OA drugs (DMOADs), like strontium ranelate and bisphosphonate.

N. Hema Rajini () Department of Computer Science and Engineering, Alagappa Chettiar Government College of Engineering and Technology, Karaikudi, Tamil Nadu, India A. Anton Smith Department of Pharmacy, Annamalai University, Chidambaram, Tamil Nadu, India © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_5

91

92

N. Hema Rajini and A. Anton Smith

Two DMOADs exhibit significant performance in animal models of OA. The cartilage and bone pathology get reduced by strontium, whereas bisphosphonates eliminate the way to remodel the bone and save its structural integrity. Recent advances in imaging approaches are regularly employed in the medicinal examination. The traditional X-ray technique which lasts for ten decades is inexpensive, faster, and simple and offers precise details regarding the modifications that take place in human bones. The present examination of OA depends mainly on the grading process of joint space narrowing (JSN) and osteophyte using an atlas. The entity grading takes more time, is highly susceptible to maximum interrater and interrater variability, and is insensitive and adequate for detecting the modifications in the joints in the earlier levels of OA. For instance, it is depicted that around 13% of cartilage quantity could be lost proper to the JSN of grade one noted, and the defection of cartilage takes place before radiography OA or knee-joint pain exist. An interesting approach to resolve this issue is the use of fractal investigation of the trabecular bone (TB) texture radiography images. It is due to the following factors: (i) TB modifies at the earlier level of OA, (ii) TB offers different characteristics of fractals, (iii) TB texture images are chosen on 2D radiographs, which are straightly relevant to the primary 3D-bone arrangement, and (iv) OA modifications are identified on the TB texture images by the use of fractal approaches. In the recent decade, several studies are aggressively functioning for devising an accurate classifier model for classification the precise classifier model for classifying provided X-ray images for suitable class based on the severity of knee OA [6]. Navale et al. proposed block-based texture examination method for the identification, and support vector machine (SVM) classification model is adopted for image classification. A set of nine evenly sized blocks of solitary input images are assumed for extracting the various texture characteristics. A total of 36 features undergo extraction and storage in feature vectors for achieving the needed classifier rate. The presented model exhibited a better classification rate while identifying normal and infected images correspondingly. Antony et al. presented a deep convolutional neural network (CNN) dependent classifier model for the automatic quantification of the severity level of OA from radiographic images [7]. It is noted that a maximum of 95.2% accuracy on the OA Initiative (OAI) database was attained from the use of around 5000 patients by the use of the linear SVM model. Anifah et al. introduced a classifier model depending upon a self-organizing map. CLAHE is applied to increase the quality of the images. A decision is made to choose if the image comes under the left/right knee depending upon the template matching model. For the identification of the cavity regions, knee image segmentation takes place by the use of Gabor kernel, row sum graph, template matching, and gray level center of the mass technique. Gray level cooccurrence matrix (GLCM) characteristics are employed to again classify the data into five OA grades. It is observed that the detection rate of the classifier model under five kinds of OA is higher compared to the images with grades 1–3. Knee X-ray image-based OA model is introduced based on the region proposal and the object classifier model based on NN is used for the identification of the knee-joint region and classification of images [9]. The presented model indicates the

Osteoarthritis Detection and Classification in Knee X-Ray Images Using. . .

93

maximum classification of OA. A volumetric-CNN-based knee cartilage model for segmenting and classifying models is presented [10]. The presented model exhibits maximum accuracy on the tested MICCAI SKI10 dataset. From the diverse preprocessing models, it is employed to enhance the contrast and also to enhance the dynamic ranges of input images [11, 12]. Diverse segmentation techniques are applied for extracting the synovial cavity regions of the input image. Finally, the classification model is employed for the segregation of the input to proper classes for the assignment of grades to the images. In this chapter, a new particle swarm optimization (PSO) model with deep neural network (DNN) named PSO-DNN technique for the identification and categorization of OA from the knee X-ray images is presented. The presented method helps to distinguish between well and affected knee X-ray images. Here, a guided filter (GF) and adaptive histogram equalization models are correspondingly employed to remove noises and enhance the images. Global thresholding-based segmentation model is employed. In addition, the images undergo classification by the use of the PSO-DNN model. For drawing a good validation, the experimentation takes place on the real-time patient-oriented images gathered from the medical organizations. From the simulation outcome, the presented PSO-DNN model confirmed the superior performance of the applied images.

2 Proposed Method In this study, a new PSO-DNN model is developed for OA classification in the IoHT environment. Initially, the IoHT devices capture the knee X-ray images and are sent to the cloud server for further process. In addition, GF-based preprocessing, global thresholding-based segmentation, and PSO-DNN-based classification processes are carried out. The detailed working of these modules is offered in the succeeding sections.

2.1 Preprocessing Initially, the X-ray images are gathered from the patients in the DICOM format and conversion takes place to grayscale images for extra computation. In the process of forming X-ray images, the images get affected by the noise. The GF is utilized to eradicate the noise. A set of three major reasons exist to adapt the GF to remove noises including (i) it is highly time efficient over the linear one, (ii) it does selective filtering, and (iii) the quantity of smoothing depends upon the quantity of local noise. At the next level, the contrast level will be enhanced. The CLAHE model is applied to increase the adaptive ranges of the images using histogram equalization [13]. Here, the images undergo partitioning into fewer portions to enhance the intensity value of the specific pixel. The bi-linear

94

N. Hema Rajini and A. Anton Smith

interpolation model is applied for joining the improved regions. The block size of 3 × 3 is applied for local processing while the parameter clip limit is kept at 0.01 to balance the level of contrast. Once the noise is removed and contrast is enhanced, extraction of synovial cavity regions takes place by the use of a segmentation model. Here, the global thresholding-based segmentation model is employed for the extraction of synovial cavity regions of the improved image.

2.2 Extracting Boundaries The main intention of the boundary extraction process is the determination of the boundaries of the cavity portions in the segmented images. A cavity is generally existing between the gap of two bones namely the femur as well as the tibia. At the same time, it is possible to generate the boundary for the false objects which lie exterior to the regions of the synovial cavity. Generally, the size of the edges is small. Here, wrong contours are eliminated and the rest of them are taken for extra computation. For the removal of a few unnecessary small object regions, an open morphological procedure is carried out. In addition, for tracing the precise boundary locations, simpler morphological boundary extraction models are utilized using a 3 × 3 structuring component. Equation (1) is applied for the boundary extraction from the segmented images. ImBoundary = Im − (Im  B)

(1)

where I and B represent the input images. Figure 1 demonstrates the different stages involved in the preprocessing stage. As depicted in Fig. 1a, the original knee X-ray image is provided. Then, the preprocessed image, i.e. noise removed image is depicted in Fig. 1b. Similarly, the segmented image by the presented model is shown in Fig. 1c.

Fig. 1 Different levels of image preprocessing (a) input image, (b) noise removed image, and (c) segmented image

Osteoarthritis Detection and Classification in Knee X-Ray Images Using. . .

95

Fig. 2 Levels of boundary extraction procedure (a) with and (b) without the existence of false object boundary

Figure 2 shows the different stages of extracting boundaries with and without the existence of false object boundaries. Figure 2a shows the boundary extraction with the existence of a false object boundary. At the same time, Fig. 2b shows the boundary extraction without the existence of a false object boundary.

3 PSO-DNN The presented model gets the input variables relevant to the question like training data which is applied and relevant to the variables of the CNN models like particles’ maximum number of layers at the time of initialization. In the PSO-DNN model, the global best particle depends upon the optimal blocks established in the swarm followed by the PSO algorithm [14]. Consequently, it is not necessary to go for manual optimization of parameters for every block. Here, the way of optimizing the parameters does not begin. The optimal blocks are sent to the subsequent generation in the form of global best particles. At present, only the particle assessment is required to restart at each iteration, however, the technique verified whether the optimal blocks are placed. Figure 3 shows the actual representation of the particles in the CNN model [15]. The presented PSO-DNN model comprises six processes namely initialization, fitness validation of every particle, a metric of the variation among the two particles, velocity determination, and particle update.

96

N. Hema Rajini and A. Anton Smith

Fig. 3 Particle representation [16]

4 Performance Evaluation For the validation of the presented technique, a set of benchmark knee X-ray images is considered. A set of normal, as well as osteoarthritis images, are present in the dataset. The sample test images are demonstrated in Fig. 4. Table 1 shows the comparison of diverse methods in terms of sensitivity, specificity, and accuracy. As shown in Fig. 5, it is obvious that the PSO-DNN technique attained the highest sensitivity of 89.09. At the same time, the ANN technique achieved a slightly reduced classification outcome with a sensitivity of 88.76. Likewise, the SVM approach obtained moderate sensitivity of 88.32. In the same way, the AdaBoost model exhibited a worse outcome with a sensitivity of 87.56. Simultaneously, the DT technique showed ineffectual results with the lowest sensitivity of 87.

Osteoarthritis Detection and Classification in Knee X-Ray Images Using. . .

97

Fig. 4 Sample knee X-ray images of (a) Normal and (b) Abnormal Table 1 Dataset details

Methods Proposed ANN [17] SVM [6] AdaBoost [18] Decision tree [19]

Sensitivity 89.09 88.76 88.32 87.56 87.00

Specificity 88.97 87.98 87.65 86.43 86.00

Accuracy 89.54 88.24 87.45 87.98 87.34

Sensitivity 89.5 89

Sensitivity

88.5 88 87.5 87 86.5 86 85.5 Proposed

ANN

SVM

AdaBoost

Decision Tree

Fig. 5 Sensitivity analysis of the PSO-DNN model

As shown in Fig. 6, it is evident that the PSO-DNN technique attained maximum classification performance with the highest specificity of 88.97. At the same time, the ANN approach achieved slightly higher classification performance with a higher

98

N. Hema Rajini and A. Anton Smith

Specificity 90

Specificity

89 88 87 86 85 84 Proposed

ANN

SVM

AdaBoost

Decision Tree

Fig. 6 Specificity analysis of the PSO-DNN model

specificity of 87.98. Likewise, the SVM model obtained moderate specificity of 87.65. Moreover, the AdaBoost model exhibited poor classification performance with a specificity of 86.43. Simultaneously, the DT approach showed an ineffectual outcome with the lowest specificity of 86. As depicted in Fig. 7, it is obvious that the PSO-DNN technique attained the highest sensitivity of 89.54. At the same time, the ANN technique achieved a slightly reduced classification outcome with a sensitivity of 88.24. Likewise, the SVM approach obtained moderate sensitivity of 87.45. In the same way, the AdaBoost technique exhibited a worse outcome with a sensitivity of 87.98. Simultaneously, the DT technique has shown ineffectual results with the lowest sensitivity of 87.34. The overall results pointed out that the presented model showed maximum outcome with an increased sensitivity of 89.09, specificity value of 88.97, and accuracy of 89.54. These experimental values pointed out that the presented model is a superior method over the compared methods. Finally, a detailed working of the XAI-enabled IoHT model is shown in Fig. 8. The figure stated that the input image is fed into the IoHT system and the model gets trained using the XAI models. Finally, the decision is notified to the physicians/users for corresponding actions. The utilization of XAI models for IoHT models assists to improve the diagnostic performance in terms of sensitivity, specificity, and accuracy.

Osteoarthritis Detection and Classification in Knee X-Ray Images Using. . .

99

Accuracy

90 89,5

Accuracy

89 88,5 88 87,5 87 86,5 86 Proposed

ANN

SVM

AdaBoost

Decision Tree

Fig. 7 Accuracy analysis of the PSO-DNN model

Fig. 8 Block diagram of IoHT model with XAI

5 Conclusion In this chapter, a new PSO-DNN model is presented for the identification and categorization of OA from the knee X-ray images. The presented method helps to distinguish between well and injured knee X-ray images. In addition, the images undergo classification by the use of the PSO-DNN model. Here, GF and adaptive histogram equalization models are correspondingly employed to remove noises and

100

N. Hema Rajini and A. Anton Smith

enhance the images. Global thresholding-based segmentation model is also utilized. The overall results pointed out that the presented model showed a maximum outcome with an increased accuracy of 89.54. These experimental values pointed out that the presented model is a superior method over the compared methods.

References 1. Zhou, X., Liang, W., Kevin, I., Wang, K., Wang, H., Yang, L. T., & Jin, Q. (2020). Deeplearning-enhanced human activity recognition for internet of healthcare things. IEEE Internet of Things Journal, 7(7), 6429–6438. 2. Jesmin, S., Kaiser, M. S., & Mahmud, M. (2020). Artificial and internet of healthcare things based Alzheimer care during COVID 19. In International conference on brain informatics (pp. 263–274). Springer. 3. Kaur, H., Atif, M., & Chauhan, R. (2020). An internet of healthcare things (IoHT)-based healthcare monitoring system. In Advances in intelligent computing and communication (pp. 475–482). Springer. 4. Arrieta, A. B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., García, S., Gil-López, S., Molina, D., Benjamins, R., & Chatila, R. (2020). Explainable artificial intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion, 58, 82–115. 5. Anifah, L., Purnomo, M., Mengko, T., & Purnama, I. (2018). Osteoarthritis severity determination using self organizingmap based Gabor kernel. IOP Conference Series: Materials Science and Engineering, 306, 12071. 6. Navale, D. I., Hegadi, R. S., & Mendgudli, N. (2015). Block based texture analysis approach for knee osteoarthritis identification using SVM. In IEEE international WIE conference on electrical and computer engineering (pp. 338–341). IEEE. 7. Antony, J., McGuinness, K., O’Connor, N. E., & Moran, K. (2016). Quantifying radiographic knee osteoarthritis severity using deep convolutional neural networks. In 23rd international conference on pattern recognition (pp. 1195–1200). IEEE. 8. Anifah, L., Purnama, I. K. E., Hariadi, M., & Purnomo, M. H. (2013). Osteoarthritis classification using self organizing map based on Gabor kernel and contrast-limited adaptive histogram equalization. Open Biomedical Engineering Journal, 7, 18–28. 9. Suresha, S., Kidzi’nski, L., Halilaj, E., Gold, G., & Delp, S. (2018). Automated staging of knee osteoarthritis severity using deep neural networks. Osteoarthritis and Cartilage, 26(1), S440–S441. 10. Raj, A., Vishwanathan, S., Ajani, B., Krishnan, K., & Agarwal, H. (2018). Automatic knee cartilage segmentation using fully volumetric convolutional neural networks for evaluation of osteoarthritis. In IEEE 15th international symposium on biomedical imaging (pp. 851–854). IEEE. 11. Ruikar, D. D., Santosh, K. C., & Hegadi, R. S. (2019). Automated fractured bone segmentation and labeling from CT images. Journal of Medical Systems, 43(3), 60. 12. Ruikar, D. D., Santosh, K. C., & Hegadi, R. S. (2019). Chapter 7: Segmentation and analysis of CT images for bone fracture detection and labeling. In Medical imaging: Artificial intelligence, image recognition, and machine learning techniques. CRC Press. 13. Ningsih, D. R. (2020). Improving retinal image quality using the contrast stretching, histogram equalization, and CLAHE methods with median filters. International Journal of Image, Graphics and Signal Processing, 12(2), 30. 14. Zhang, X., Liu, H., & Tu, L. (2020). A modified particle swarm optimization for multimodal multi-objective optimization. Engineering Applications of Artificial Intelligence, 95, 103905.

Osteoarthritis Detection and Classification in Knee X-Ray Images Using. . .

101

15. Gadekallu, T. R., Khare, N., Bhattacharya, S., Singh, S., Maddikunta, P. K. R., & Srivastava, G. (2020). Deep neural networks to predict diabetic retinopathy. Journal of Ambient Intelligence and Humanized Computing, 1–14. https://doi.org/10.1007/s12652-020-01963-7 16. Fernandes, F. E., Jr., & Yen, G. G. (2019). Particle swarm optimization of deep neural networks architectures for image classification. Swarm and Evolutionary Computation, 49, 62–74. 17. Hegadi, R. S., Navale, D. I., Pawar, T. D., & Ruikar, D. D. (2019). Osteoarthritis detection and classification from knee X-ray images based on artificial neural network. In Recent Trends in image processing and pattern recognition: Second international conference, RTIP2R 2018, Solapur, India, December 21–22, 2018, Revised Selected Papers, Part II 2 (pp. 97–105). Springer. 18. Saad, G., Khadour, A., & Kanafani, Q. (2016). ANN and Adaboost application for automatic detection of microcalcifications in breast cancer. The Egyptian Journal of Radiology and Nuclear Medicine, 47(4), 1803–1814. 19. Qing-Yun, S., & Fu, K. S. (1983). A method for the design of binary tree classifiers. Pattern Recognition, 16(6), 593–603.

Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques ˙ Ilhan Uysal

and Cafer Çali¸skan

1 Introduction In recent years, there has been a huge increase in the amount of biological data obtained along with the great breakthroughs in technology. Consequently, traditional approaches in data analysis have become inadequate due to the huge amount of data. Therefore, some new approaches take place to process this data more efficiently. The most frequent one is to apply machine learning techniques for facilitating the management of the data. In this sense, machine learning algorithms can potentially reveal more outcome possibilities compared to the traditional methods. For instance, it was much more difficult to detect the symptoms of a disease before; however, machine learning algorithms can capture them in a more comprehensive way provided that accurate relative data is present. Moreover, some machine learning algorithms have great potential to produce results with higher accuracy in the early stages of some health problems. Machine learning techniques are generally black-box. In other words, it is not possible to name the solution exactly. People normally have to rely on the decisions reached by black-box systems. However, this creates a contradiction. Artificial intelligence techniques that can explain the results or inferences it reaches are called Explainable Artificial Intelligence. The inference processes and results of such systems can be understood by humans. But before that, traditional machine learning

˙ Uysal () I. Burdur Mehmet Akif Ersoy University, Bucak Emin Gülmez Technical Sciences Vocational School, Computer Technology Department, Burdur, Turkey e-mail: [email protected] C. Çali¸skan Antalya Bilim University, Faculty of Engineering, Computer Engineering Department, Antalya, Turkey e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_6

103

104

˙ Uysal and C. Çali¸skan I.

has been employing interpretable aspects to ensure simple and understandable enough solutions. Additionally, the sensor-based devices that are incorporated with the Internet of Things (IoT) and their integration with mobile technologies are being referred to as the Internet of Healthcare Things (IoHT). This study aims to analyze biological data using interpretable machine learning techniques to assist doctors in medical diagnoses. Furthermore, the solutions presented in this work have the potential to contribute to Explainable Artificial Intelligence (XAI) and the Internet of Health Things (IoHT). This study is mainly structured as follows: The first section includes preliminaries, then the second section discusses some machine learning techniques in the estimation of VLDL cholesterol. The next section includes the research findings and finally, the last section discusses the results together with the contribution of the machine learning techniques in performance and precision.

2 Preliminaries Machine Learning (ML) is the use of algorithms that helps with modeling computer systems to learn from data. Nowadays, it is possible to do many jobs by using machine learning. As a field, machine learning has many benefits such as the determination of relationships within large amounts of data, an easier processing of image-based data, helping experts with difficult decisions, and more rapid processing of large amounts of data which is impossible to be done by the human brain at short notice [1]. Also, in order to ensure understandable solutions, in which the input-output connection can be interpreted, some simple machine learning techniques may be applied in specific applications. There are various software available for applying machine learning techniques. Rapidminer is one of such software that competes with many paid software in terms of usage rates and preference in order. It can also meet the needs of both beginner and expert level users. It also supports many file extensions such as csv, dat, and log. It provides more than 400 algorithms. Database systems can be with ease run with the Rapidminer. Users do not need to adjust individual steps or parameters because experimental designs are automatically optimized through the meta operators. The software is written in Java language and compatible with languages such as Python, Weka, or R [2]. In this study, the latest version of Rapidminer Studio (version 9.6) has been used for applying interpretable machine learning algorithms such as generalized linear model (GLM), decision tree, and gradient boosted trees. In what follows the basics of these techniques are explained.

Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques

105

2.1 Generalized Linear Model The generalized linear model was first discussed in 1972 by Nelder and Wedderburn. This model, which is used mostly in social sciences and medical applications, also plays an important role in the analysis of life data. This algorithm is an extension of traditional linear models as well as fits generalized linear models to the data by maximizing the log-likelihood. It is a combination of both linear and non-linear regression models that takes into account the non-normally distributed dependent variable. It is a powerful alternative to data conversion [3]. In some empirical studies, the generalized linear model is used as an alternative to data transformation in cases where the dependent variable does not conform to normal distribution [4]. A linear predictor, described in Eq. (1), along with a link and a variance function constitute the components of a generalized linear model. ni = β0 + β1X1i + · · · + βpXpi

(1)

The link function describes how the mean of the distribution depends on the linear predictor. qE (Yi ) = μi

(2)

g (μi ) = ηi

(3)

A variance function describes how the variance depends on the mean where the dispersion parameter ϕ is a constant [5]. var (Yi ) = ϕV (μ)

(4)

2.2 Decision Trees Decision trees are one of the most popular supervised machine learning algorithms used for both classification and regression tasks [6]. Decision trees are algorithms that show more understandably the information set that the classifier has and show them in a way tree by sorting in a certain arrangement the class options and the situations depending on the probabilities [7]. They can be run on categorical and numerical data. In decision trees, it is used some special terms similar to those of a real tree, such as branches and leaves. It has decision nodes and leaf nodes in a decision tree [8]. Decision nodes are attributes used to make decisions in the dataset, classify or predict while leaf nodes keep decisions. The node at the topmost of the tree is called the root node. To reach a decision, a certain path is followed from the root of the tree to the leaf nodes [9].

˙ Uysal and C. Çali¸skan I.

106

With the algorithms used in the decision tree, it is aimed to extract a decision tree from a given data set by minimizing the generalization error [10]. The algorithm selection can be made according to the type of target variables. Iterative Dichotomiser 3 (ID3), Successor of ID3 (C4.5), Classification and Regression Tree (CART), and Chi-square Automatic Interaction Detector (CHAID) are the algorithms frequently used for decision trees [11]. One of the simplest decision tree algorithms that only work with categorical attributes is ID3 by J. R. Quinlan. ID3 generates a decision tree from a dataset that is represented by a table in general. It uses a top-down, greedy search through the given columns when constructing a decision tree, where it selects the attribute that is best for the classification of a given set. To decide what attribute is the best selection in terms of constructing a decision tree, ID3 uses Entropy and Information Gain [12]. Entropy is the measure of uncertainty or randomness in a given data. Intuitively, it indicates the foreseeability of a particular event. If a result of an event has a likelihood of 100%, the value of entropy is zero and if a result is 50%, the value of entropy reaches the maximum value as it projects perfect randomness. It is with the lowest possible probability to determine the outcome, so as a result, the entropy is likely to get the highest possible value. Building a decision tree requires the calculation of two types of entropy. The first one is the entropy E(S) using the frequency table of one attribute. To give a current state S and a probability P(x) of an event x of that state S:  E(S) = −P (x) log2 P (x) (5) x∈X

Let A be a selected attribute, S a current state with this attribute A, and P(x) a probability of an event x of the same attribute A, then E(S, A) denotes the entropy that is using the frequency table of these two attributes S and A [13]. E (S, A) =



[P (x)· E(S)]

(6)

x∈X

While the Eq. (6) relates to the Entropy of an attribute A, the Eq. (5) is the Entropy of the entire set. Information gain (IG) is a criterion showing how effective it is a given feature is in classification and takes a value between 0 and 1. It is denoted by IG(S, A) for a state S is the important change in entropy after deciding on a certain attribute A. The relative change in entropy to the independent variables can be measured as shown in Eq. (7): IG (S, A) = E(S) − E (S, A)

(7)

Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques

107

Constructing a decision tree is all about selecting each attribute (A) to calculate Information Gain and finding such an attribute that returns the highest IG. The next decision node for the tree will be this attribute [14]. The C4.5 algorithm can be considered an improved version of the ID3 algorithm. The main difference between this algorithm and the ID3 algorithm is that it uses normalization. While in the ID3 algorithm, entropy or information gain is calculated and decision points are determined, in the C4.5 algorithm, entropy values are kept as a ratio [15]. CART, which is capable of working with both numerical and categorical data, uses the Gini algorithm, regression trees, and random forest algorithms as branching criteria and produces binary trees. In the CART algorithm, it is performed the partitioning process by applying a certain criterion in a node. For this, first, it is taken into account the values with all the qualities, and after all matches, two splits are obtained [16]. The CHAID algorithm was created by Kaas in 1980 when there was no statistically significant difference to calculate the best division by combining the possible category pair of the prediction variable in pairs that fit the target variable. The Chi-square test is used instead of entropy or Gini metrics used to select the most suitable sections. To calculate the best division, the prediction variables are combined until there is no statistically significant difference in a pair that fits the target variable. The main difference between CHAID and other techniques, while ID3, C4.5, and CRT derive binary trees, CHAID derives multiple trees [17].

2.3 Gradient Boosted Trees Gradient boosted trees is a machine learning technique that is used for regression and classification problems. This creates a model of decision trees, typically combined with weak prediction models [18]. The purpose of any supervised learning algorithm is to identify and minimize a loss function [19]. MSE = a0 +

 p  yi − yi 2

(8)

 p p  It is shown yi is the ith target value, .yi is the ith prediction, and L = . yi, yi is the loss function in Eq. (8).

3 The Research Findings and Discussion In this study, the dataset is obtained from the Burdur Provincial Health Directorate. It contains the blood test results of the patients at the internal-medicine-polyclinic of Burdur State Hospital from 2017 to 2018. It has 67 different laboratory analyses

108

˙ Uysal and C. Çali¸skan I.

Fig. 1 The comparison of correlation coefficient of models

Fig. 2 The comparison of root mean squared error values of models

Fig. 3 The comparison of absolute error values of models

that consist of biochemistry and hormone test results of exactly 20,004 patients. In this dataset, there are exactly 6883 male patients and 13,121 female patients [20]. The comparison of the correlation coefficient of models used is given in Fig. 1. Although the best performing model is GLM, other models also found very close to 1 (0.999 and 0.996) and a positive relationship. Root mean squared error (RMSE) measures the error of a model in predicting quantitative data. The comparison of the applied models in terms of RMSE is given in Fig. 2. According to our computation, GLM has the best performance and the value is 0.655. In comparing performances, another criterion used is the absolute error value, which shows the average vertical distance between each real value and the line that best fits the data. According to our computation, the decision tree technique has the best performance in terms of the absolute error value. The decision tree algorithm has an absolute error value equal to 0.109 as is shown in Fig. 3. The mean squared error (MSE) indicates how close a regression curve is to a series of points. MSE measures the performance of the estimator in the machine learning model and is always positive, and predictors with an MSE value close to zero are assumed to perform better. The comparison of squared error values of the applied models is given in Fig. 4. While the MSE value of the GLM algorithm is

Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques

109

Fig. 4 The comparison of squared error values of models

Fig. 5 The weights by correlation

0.441, the decision tree algorithm has an MSE value equal to 1.01, and the gradient boosted has an MSE value of 2.245. Figure 5 lists the weights of the attributes in computing the correlation. According to our computation, the most effective attribute in the estimation of VLDL cholesterol value is triglyceride. Apart from that, other effective ones are total cholesterol, HDL cholesterol, glucose, LDL cholesterol, ALT, free T4, insulin, potassium, and chlorine.

3.1 Generalized Linear Model In this study, it is observed that attributes are supporting the prediction and attributes contradict the prediction. In Fig. 6, the list of supporting and contradicting attributes that are statistically significant in predicting the VLDL cholesterol values is given with their significance levels. According to this list, the significant supporting attributes are triglyceride, lipase, creatinine, and free PSA in the reference range by the GLM. The attributes of CEA, chlorine, and globulin are contradicting factors in predicting the values.

110

˙ Uysal and C. Çali¸skan I.

Fig. 6 Important factors for normal values of VLDL cholesterol by the generalized linear model

The predictions chart below shows the predictions (by using the generalized linear model) versus the actual values for the 40% validation cases. Each plot in the graph represents a specific prediction. As the plots get closer to the orange line, the model gets better [21]. See Fig. 7.

3.2 Decision Tree With the assumption that VLDL cholesterol values are within the reference range, applying the Decision Tree technique determines the significant supporting attributes as triglyceride, lipase, creatinine, and free PSA. The attributes of CEA, chlorine, and globulin are contradicting factors in prediction. See Fig. 8. The predictions chart obtained for VLDL cholesterol values by applying the decision tree algorithm is given in Fig. 9. Optimal Parameters show the model’s performance for different parameter settings. Error rates for maximal depth in normal values of VLDL cholesterol by applying the decision tree algorithm are given in Fig. 10. Accordingly, when the maximal depth is at least 10, the error rate drops to 0.6%.

3.3 Gradient Boosted Trees With the assumption that VLDL cholesterol values are within the reference range, applying the gradient boosted trees technique determines the significant supporting attributes as triglyceride, lipase, creatinine, and free PSA. The attributes of CEA, chlorine, and globulin are contradicting factors in prediction. See Fig. 11.

Fig. 7 Predictions Chart of VLDL cholesterol by the generalized linear model

Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques

111

112

˙ Uysal and C. Çali¸skan I.

Fig. 8 Important factors for normal values of VLDL cholesterol by the decision tree

The predictions chart obtained for VLDL cholesterol values by applying the gradient boosted trees algorithm is given in Fig. 12. Error rates for parameters in normal values of VLDL cholesterol by applying the gradient boosted trees algorithm are given in Fig. 13. Accordingly, the number of trees has been found as 90, maximal depth 7, learning rate 0.100, and error rate 1.6%.

4 Conclusion Doctors request blood tests according to the complaints of patients. For some particular diseases or health problems, certain test results are requested by doctors to find out about the underlying health issues. In this manner, the test results are crucial for them. Although the interconnection between the test results and health issues is sometimes complicated, some studies aim to help doctors to resolve such connections. In this study, it is observed that age and gender are important factors in the prediction of all test results, as a result, they are also significant for VLDL cholesterol. In addition, other important factors in predicting VLDL cholesterol are

Fig. 9 Predictions chart of VLDL cholesterol by the decision tree

Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques

113

˙ Uysal and C. Çali¸skan I.

114

Error Rate

40%

20%

0% 5

10

15 Maximal Depth

20

25

Fig. 10 Error rates for maximal depth

Fig. 11 Important factors for normal values of VLDL cholesterol by the gradient boosted trees

triglycerides, lipase, creatinine, and free PSA. This outcome is verified with various techniques such as GLM, decision tree, and GBT algorithms in this study. They all have high success rates which are 95% or higher. For details see Table 1. As the solutions reached in this study have the potential of XAI and IoHT, this topic will be a source of inspiration for future studies. Thanks to XAI, patients will know how the solutions given to them are realized, and thanks to IHoT, they will be able to communicate with medical facilities without interruption. In this way, there will be real-time information exchange between patients and medical facilities.

Fig. 12 Predictions chart of VLDL cholesterol by the gradient boosted trees

Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques

115

˙ Uysal and C. Çali¸skan I.

116

Fig. 13 Error rates for parameters in the gradient boosted trees Table 1 Comparison of algorithms’ performances Algorithm GLM DT GBT

Correlation 1 0.999 0.996

RMSE 0.655 0.8 1.455

Absolute error 0.498 0.109 0.324

Squared error 0.441 1.01 2.245

References 1. Turgut, S. (2017). Cancer diagnosis using machine learning methods. MSc thesis, Institute of Graduate Studies in Science and Engineering Department of Computer Engineering, Istanbul, Turkey, pp. 19–20. 2. Uysal, I., Bilen, M., & Ulukus, S. (2018). Analysis of classification algorithms with rapidminer (pp. 517–521). Gece Kitapligi Publishing House. 3. https://docs.rapidminer.com. Access date: 08.05.2021. 4. Karadag, O., & Aktas, S. (2018). Generalized estimating equations for genetic association studies of multi-correlated longitudinal family data. Gazi University Journal of Science, 33(1), 273–280. 5. Shrestha, N. (2019). Estimating the probability of earthquake occurrence and return period using generalized linear models. Journal of Geoscience and Environment Protection, 7, 11– 24. 6. Halima, E., Zahra, B., & Hassan, A. (2018). A comparative study of algorithms constructing decision trees. In Proceedings of the international conference on learning and optimization algorithms: Theory and applications – LOPAL’18 (pp. 1–5). Association for Computing Machinery. 7. Danandeh Mehr, A., Bagheri, F., & Safari, M. (2020). Electrical energy demand prediction: A comparison between genetic programming and decision tree. Gazi University Journal of Science, 33(1), 62–72. 8. Quinlan, J. R. (1986). Induction of decision trees. Machine Learning, 1, 81–106. 9. Bilgin, M., & Yılmaz, A. (2018). Machine learning (pp. 90–107). PapatyaBilim Publishing. 10. Balaban, M. E., & Kartal, E. (2018). Data mining and machine learning (2nd ed., pp. 90–91). Caglayan BookStore.

Prediction of VLDL Cholesterol Value with Interpretable Machine Learning Techniques

117

11. Chien-Liang, L., & Ching-Lung, F. (2019). Evaluation of CART, CHAID, and QUEST algorithms: A case study of construction defects in Taiwan. Journal of Asian Architecture and Building Engineering, 18(6), 539–553. 12. Sudrajat, R., Irianingsih, I., & Krisnawan, D. (2017). Analysis of data mining classification by comparison of C4.5 and ID algorithms. IOP Conference Series: Materials Science and Engineering, 166, 012031. 13. https://www.thelearningmachine.ai/. Access date: 08.05.2021. 14. Wang, Y., Li, Y., Song, Y., Rong, X., & Zhang, S. (2017). Improvement of ID3 algorithm based on simplified information entropy and coordination degree. Algorithms, 10(4), 123–124. 15. Ture, M., Tokatli, F., & Kurt, I. (2009). Using KaplanMeirer analysis together with decision tree methods (CART, CHAID, QUEST, C4.5, and ID3) in determining recurrence-free survival of breast cancer patients. Expert Systems with Applications, 36, 2017–2026. https://doi.org/ 10.1016/j.eswa.2007.12.002 16. Yang, Y., Velayudhan, A., Thornhill, N. F., & Farid, S. S. (2015). Manufacturability indices for high-concentration monoclonal antibody formulations. Computer Aided Chemical Engineering, 37, 2147–2152. 17. Ozkan, Y., & Erol, C. S. (2017). Bioinformatics DNA microarray data mining (2nd ed., pp. 227–282). PapatyaBilim University Publishing. 18. Tapodhir, A., Saurav, C., Suman, N., Abdul, M. C., & Sanjeev, K. (2019). Symptoms to disease mapping and doctor recommendation system. International Journal of Engineering and Advanced Technology (IJEAT), 9(1), 2249–8958. 19. Natekin, A., & Knoll, A. (2013). Gradient boosting machines, a tutorial. Frontiers in Neurorobotics, 7, 20–21. ˙ & Çalı¸skan, C. (2019). Some machine learning techniques for medical diagnosis. 20. Uysal, I., Master thesis, Antalya Bilim University Institute of Science, p. 11. 21. https://docs.rapidminer.com/8.1/studio/operators/validation/visual/lift_chart.html.Access date: 11.05.2021.

A Survey of Interpretable Cognitive IoT Devices for Personal Healthcare Rav Raushan Kumar Chaudhary and Kakali Chatterjee

1 Introduction E-healthcare refers to the use of the Internet to monitor healthcare. Internet of Things (IoT) is gaining wide acceptance and is increasingly adopted in many aspects of our daily lives [1]. The IoT technology provides an effective structured method to improve the health and well-being of people. IoT-based systems are expected to reshape the healthcare industry in terms of social benefits and penetration rates and profitability. Designing standalone portable devices is no longer sufficient, but it is essential to create a complete ecosystem, where sensors on the body area network seamlessly synchronize data with the cloud service through the infrastructure of the IoT [2]. There are three major components in medical Internet of Things (HealthIoT) systems such as (i) a body area sensor network, (ii) a gateway connected to the Internet, and (iii) cloud and big data carriers. The medical sensor and actuator network captures biomedical and environmental signals through ubiquitous recognition, detection, and communication functions from the body and the room. The data is then sent to the gateway using a wired or wireless communication channel (such as Bluetooth, WiFi, ZigBee, or 6LoW- PAN) [3]. This data is sent to several geographically dispersed smart e-Health gateways before being stored in the cloud for various reasons such as statistical and epidemiological medical research [4]. In today’s COVID-19 scenario, the cognitive internet of medical things (CIoMT) has been used for real-time tracking of cured patients, death records, and patients under treatment. CIoMT also helps doctors to monitor patient conditions by observing body parameters such as pulse rate, temperature, and oxygen level.

R. R. K. Chaudhary () · K. Chatterjee Department of Computer Science and Engineering, National Institute of Technology Patna, Patna, India e-mail: [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_7

119

120

R. R. K. Chaudhary and K. Chatterjee

Security and privacy are major concerns for patient health data. Theft and loss, insider misuse, and inadvertent action are all hazards to the healthcare industry. Designing an efficient Internet of Things-based system in healthcare is a difficult undertaking due to the following major challenges. First, the sensor network technology used must be resource efficient and adaptable to electronic medical applications. Medical sensor nodes, particularly implanted nodes, have a much lower processing power, memory, and power delivery than the sensors in other sensor network domains. Secondly, a considerable amount of energy is dissipated during transmission. The transmission of electrocardiogram (ECG) signals requires a bandwidth of 4 kbps per channel. Thus suitable IoT devices are essential in the healthcare monitoring process. The contributions of this chapter are as follows: • A survey of sensor devices with web technologies used in healthcare monitoring systems has been explained. • It also provides issues and challenges of personal healthcare system. The remaining parts of this paper are structured as follows: Section 2 discusses related work, and Sect. 3 discusses IoT-based personal healthcare systems. In Sect. 4, the development of smart personal healthcare work has been described. Section 5 describes the issues and challenges of IoT-based e-healthcare. Finally, the conclusion of the chapter is discussed in Sect. 6.

2 Related Work Several techniques to solve the security problems in an IoT-based e-healthcare system have been discovered. The fundamental goal of an IoT-based e-healthcare system is to gather health data from various users and share it with them in order to deliver various healthcare-related services. Table 1 shows the various technologies of IoT-based healthcare systems.

3 IoT-Based Personal Healthcare 3.1 Smart IoT Device Smart IoT devices generally have low power, low memory, and low processing speed. All IoT devices are controlled by a microcontroller. Memory can be found in RAM or ROM. The file sizes range from 1MB to 1GB. Two companies, ARM and Intel, develop the processor. Some processors range from 8 to 16 bits. Power is in short supply. The devices should be connected to the Internet. Power management, communication stack, firmware upgrade, processing, and customizable security

A Survey of Interpretable Cognitive IoT Devices for Personal Healthcare

121

Table 1 Relevant literature on various technologies of IoT-based healthcare system S.no 1.

References Alaba et al. [5]

Year 2017

Details and domain A survey has been performed with IoT security

2.

Buzzannca et al. [6]

2018

3.

Wu et al. [4]

2018

Establish the trust relationship between the communicating parties in the different application environments Describe a sliding window and Dempster- Shafer evidence theory-based trust model for the cloud users

4.

Pasha et al. [7]

2017

IoT security, RFID sensor

5.

Dhanvijay et al. [8] Islam et al. [9]

2020

7.

Tamilsevi et al. [10]

2020

WSN IoT-based E-healthcare Smart healthcare monitoring sys tem using IoT with different conditions IoT-based healthcare system testing with Arduino uno

8.

Archarya et al. [11]

2020

6.

2020

IoT-EHS environment with Raspberry Pi, heart rate sensor, ECG sensor, respiration and b.p sensor

Objective and functionality The taxonomy of the current security threats in contexts of application, and architecture is presented Dynamic trust-based access control is used

The sliding windows will reflect the well-timed first-hand interaction evidence, whereas the Dempster-Shafer evidence interaction theory is used for calculation of the direct trust between two entities Sensor, MQTT, CoAP technology are used for healthcare Security Real-time patient data analysis for different conditions Measure heart rate, blood pressure, eye movement, and oxygen level using sensors Real-time health monitoring

Limitation Only explore the threats, provide no security solution

Limited trust level in cryptographic server

The model does not show how the credibility is added to the entity profile

Lack of efficiency is found

Only consider security issues Suffers from data redundancy and network congestion Just deals with sensor details.

Increase with high response time

122

R. R. K. Chaudhary and K. Chatterjee

features are all required for IoT devices. On the chip, there is an ARM architectural system (SOCs) which has 128 KB of RAM and 1 MB of storage.

3.2 Smart Body Sensors Body sensors are used for data capturing. There are the following sensors:1. LM35 Sensors – The body temperature sensor (LM35) series measures the correct temperature in Centigrade. There are circuits that have a voltage output. The LM35 has an advantage over the linear temperature sensor from Kelvin. The LM35 temperature sensor can be used to measure a patient’s body temperature. This sensor produces a degree Celsius analog output. The voltage is determined by the microcontroller through the use of an analog-to-digital converter (ADC). When the sensor temperature device is turned on, the microcontroller can initialize ADC, and the output of the sensors is applied to the ADC’s input channel, and the imported SMS is saved in the microcontroller’s memory. Each sensor node has data to collect. Important health data has been sent to a central hub in the cloud via Wi-Fi. A GSM module is included in the hardware model. It can be used as a port for wireless data transfer communication with the microcontroller. The LM35 is a centigrade sensor that measures temperature in ◦ C. Its temperature sensor is a precision integrated circuit. It employs an electrical output to measure temperature. External calibration is not required for the LM35. The LM35 has a temperature range of 0.5 ◦ C at room temperature and 0.7 ◦ C from 40 ◦ C to +150 ◦ C. It is shown in Fig. 1. 2. DHT11 Sensors – The room temperature sensor (DHT11) is one of the important temperature sensors. 3. MQ-9 Sensors – The CO sensor (MQ-9) can detect CH4, CO, and LPG. The MQ9 sensor has high sensitivity and quick response time. Hardware components used in healthcare include ESPO2, MQ-9, MQ-135, heartbeat sensor, and DHT11. All sensors are connected to EPSO2 (MQ-9, MQ-135, heartbeat sensor, and DHT-11). The DHT-11 sensor can detect the ambient temperature. The LPG, CO, and CH4 values are discovered using the MQ-9 sensor. 4. MQ-135 Sensors – The MQ-135 sensor detects NH3, nicotine, benzene, smoke, and CO2. The ESP2 contains a built-in Wi-Fi module. DHT-11 is wired to the D14 pin. The heartbeat sensor can be connected to the ESP32’s D26 pin. MQ-9 connects to ESP32’s D27. MQ-135 connects to ESP32’s D34. 5. ECG Sensors – The procedure of recording the electrical activity of the heart over a period of time is known as electrocardiography (ECG). Electrodes are attached to the body’s skin. Each cardiac beat’s electrophysiologic patterns are depolarized, and the ECG reveals whether the heart is normal. The Bluetooth module can be used to create a wireless headset. The GSM module can be used to improve the health of patients. It is employed in a patient monitoring system

A Survey of Interpretable Cognitive IoT Devices for Personal Healthcare

123

Fig. 1 Body sensor device

that is intelligent. The GSM module has been used to communicate with the hospital’s specific departments as well as other locations on the patient. It can send out alert messages based on patient health parameters, including heart rate, ECG, and blood pressure. 6. Heart Rate Sensors – The heart rate sensor is a critical health parameter where an optical sensor captures the changes in the volume of blood vessels during the heart pump. 7. SPD015GDN Sensors – A tiny piezoresistive pressure sensor (SPD015GDN) can be interfaced with a microcontroller to monitor blood pressure in rages. Doctors are permitted to use the one-of-a-kind patient-identification system. The patient-identification device could support any biometrics, fingerprint, or form of biometrics and voice recognition. The medical information contained in a biometric scan is that of the patient. Age, gender, general health, and family history are examples of patient medical information.

124

R. R. K. Chaudhary and K. Chatterjee

3.3 Web-Technologies for IoT System There are existing web technologies that can be used to develop IoT systems. Those protocols are HTTP RESTful, CoAP, MQTT, XMPP, etc. HTTP RESTful uses a TCP transport mechanism for smart home and smart grid applications. Protocol – web methods There are other message methods available, including get(), request(), response(), and HTTP request(). 1. HTTP – HTTP is a client-server protocol that is based on connections. The request() method is available in HTTP. In the get() method, the HTTP request message is applied. HTTP has shown to be a reliable transmission control protocol. TCP has a two-way and three-way handshake. 2. CoAP – A service mechanism is included in the CoAP. In an internet context, a COAP connection is made between an IoT device and an IoT gateway. The CoAP environment and the HTTP proxy are in communication. In IoT devices, the CoAP HTTP model proved effective. The get() method of a CoAP request is possible. 3. MQTT – MQTT is a Request/Response-based protocol used in remote monitoring and controlling of devices. Another protocol XMPP is used for remote management of major appliances which is also based on TCP connection.

3.4 IoT Smart Healthcare Components There healthcare components as shown in Fig. 2 are as follows: 1. Data collector – The data collector is just like a sensor device that collects or monitors data. It converts the health state to a digital signal. 2. IoT Gateway – With the help of the Internet protocols IPV4 and IPV6, an IoT gateway connects a local processing unit, a distant monitor, and temporary storage. An IoT gateway provides an internet protocol converter. 3. Backend Facilitator – The backend facilitator is like just a third-party service. These tools provide security management and application development. 4. IoT Application – Access application is an IoT system service that is used to access IoT applications. SHU = f(DC, iGW, Bf,AA), where SHU stands for smart health unit, DC for data collector, iGW for IoT gateway, Bf for backend facilitator, and AA for access application. f = DC (data collection, data forwarding) CT = f(iGW), where CT = f (Bluetooth, WiFi, Zwave) The letter CT stands for communication technology. IoT framework is shown in Fig 2. In this figure, three layers have been used. They are explained below 5. Sensor Layer – In an IoT-based e-healthcare system, the sensor layer is the first layer of address resources. Communication technology, bar code with RFID

A Survey of Interpretable Cognitive IoT Devices for Personal Healthcare

125

Fig. 2 Healthcare monitoring framework

technology, data collectors such as sensor devices, and IoT gateway devices are all part of the sensor layer (IGW). 6. Network Layer – The second place goes to the network layer of an IoT-based e-health care system. It acts as a backend connectivity facilitator and an IoT gateway. Communication technologies, backend cloud services, IoT gateway, a list of registered remote consultants, healthcare solutions via smart phones, and so on are all examples of network sensor components. 7. Service Access Layer – The smart health unit’s service is given to the service access layer via an access application. Medical professionals, mana gement authorities, smart computing devices and desktop computers, smart applications and new web technologies, and so on are all examples of service access layers.

126

R. R. K. Chaudhary and K. Chatterjee

4 Development of Smart Personal Healthcare Healthcare architecture is divided into two parts: the first is computer hardware, which can be described, and the second is an artificial neural network technology, which can be defined using fuzzy logic. Sensor, Zigbee, RFIG, Microcontroller, ADC, GSM Module, and Graphics User Interface are examples of computer hardware components. Machine learning, ANN (Artificial Neural Networks), and other artificial intelligence technologies are examples. In a safe health monitoring system, a neural network defines fuzzy logic. MBLAB employs a virtual sensor connection to find simulation results for a person pro. Person pro employs a virtual sensor connector to find simulation outcomes. Authentication, trust, and access control are all security requirements here. 1. Remote Monitoring System – As we know, IoT plays a very important role in many applications. There are three phases such as clinical, remote monitoring, and context awareness. During data collection, there can be risk of human monitor error but in such an automatic device monitoring system, the chance gets low. Also, e-healthcare can be able to save money as well as good quality health monitoring. But it could have threats and attacks. If data transmits between the healthcare server and user. Data information can be hacked by hackers. Clinical care plays a very important role for every patient, whereas patient has health monitored by the doctor, nurse, IoT sensor finder (physiological device), and health monitoring machine. IoT is the best source of health monitoring because it gives the alert message as well as time to time updating. 2. Clinical care – Clinical care is extremely vital for every patient who is monitored by a doctor, nurse, IoT sensor finder (physiological gadget), and health monitoring system. Because it provides alert messages as well as real-time updates, IoT is the ideal source of health monitoring. 3. Healthcare Services – IoT healthcare services are critical components of an IoTbased e-healthcare system. It has a wide range of advantages. AAL (Ambient Assistive Living), M-IOT (The Internet of m-health things), ADR (Adverse Drug Reaction), CHI (Children health information), WDA (We- arable device access), SMA, IEH (Indirect Emergency Healthcare), EGC (Emergency Gateway Configuration), and so on are examples of healthcare services. However, new healthcare research has revealed the potential of IoT in the healthcare sector. (18) Assistive living services for the elderly are available. It can be used for RFID and other purposes. IoT healthcare AAL service architecture can provide security, control, and communication. 4. Healthcare Application – The application of WBAN-based IoT in e-healthcare is extensive. There are many different types of IoT services, such as linked cars, connected healthcare, and intelligent grid. The applications of IoT healthcare are subdivided into single-condition and clustered condition. Glucose-level-sensing, electrocardiogram monitoring, blood pressure monitoring, body temperature monitoring, and oxygen saturation monitoring are all examples of healthcare applications that can be used in a single circumstance. Cluster condition

A Survey of Interpretable Cognitive IoT Devices for Personal Healthcare

127

healthcare applications include rehabilitation systems, medication administration, wheelchair management, immediate healthcare solutions, and healthcare solutions via cell phones.

5 Issues and Challenges of IoT-Based E-Healthcare Security and privacy are the most commonly debated aspects of the EHS among both businessmen and healthcare researchers. We’ve covered a lot of EHS security and privacy issues in this area. The subsections that follow provide more information on each topic. 1. Healthcare data and service issues – In the EHS, healthcare data plays a critical role. When healthcare consumers put their data in a remote storage site, they are always concerned about losing the data or losing the data’s privacy. As a result, the data owner must guarantee that all security services and strategies are correctly applied before keeping their data. 2. Trust Issues – In the EHS, trust management is particularly crucial, especially when the EHS has a distributed design. It was designed to work with a variety of security mechanisms in order to improve the security of patient-sensitive data. 3. Anonymity – The patient’s medical records are a key element of the EHS. As a result, patient information security is a major concern in the EHS. The patient’s info can be used for two purposes: main and secondary. 4. Access control – In the EHS, access control is the most commonly researched topic. It is a strategy for identifying the access choice in every access request. This decision to grant access is based on a set of access rules and policies. As a result, it aids in the protection of critical healthcare data and resources from unauthorized users. 5. Authentication Issues – The leakage of user credentials from credential storage is used to investigate authentication vulnerabilities. For the most part, EHS uses a user name and password to authenticate users. The web interface provides access to all EHS data and services. 6. Compliance and legal issues – The Service Level Agreement (SLA) is a contract that contains an agreement between two communicating parties regarding services. Both parties agree to the terms and conditions of the services and sign them. The deployed cyber laws, forensics techniques, and activities in the EHS do not fully protect the security and privacy of patient data.

6 Conclusion The current study presents a specialized framework for providing smart health care in developing countries, particularly in rural areas. The framework investi

128

R. R. K. Chaudhary and K. Chatterjee

gates a variety of features of IoT technology for smart health services, including interoperability and standards difficulties, limited Internet settings, and specialized applications. This study also focused on the challenges of the modern healthcare system.

References 1. Xue, K., Ma, C., Hong, P., & Ding, R. (2013). A temporal-credential- based mutual authentication and key agreement scheme for wireless sensor networks. Journal of Network and Computer Applications, 36(1), 316–323. 2. Abie, H., & Balasingham, I. (2012). Risk-based adaptive security for smart IoT in ehealth. In Proceedings of the 7th international conference on body area networks, BodyNets’12 (pp. 269–275). ICST (Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering). 3. Binu, P. K., Thomas, K., & Varghese, N. P. (2017). Highly secure and efficient architectural model for iot based health care systems. In 2017 international conference on advances in computing, communications and informatics (ICACCI) (pp. 487–493). IEEE. 4. Wu, F., Xu, L., Kumari, S., Li, X., Das, A. K., & Shen, J. (2018). A lightweight and anonymous RFID tag authentication protocol with cloud assistance for e-healthcare applications. Journal of Ambient Intelligence and Humanized Computing, 9(4), 919–930. 5. Alaba, F. A., Othman, M., Hashem, I. A. T., & Alotaibi, F. (2017). Internet of things security: A survey. Journal of Network and Computer Applications, 88, 10–28. 6. Buzzanca, M., Carchiolo, V., Longheu, A., Malgeri, M., & Mangioni, G. (2017). Direct trust assignment using social reputation and aging. Journal of Ambient Intelligence and Humanized Computing, 8(2), 167–175. 7. Pasha, M., & Shah, S. M. W. (2018). Framework for e-health systems in iot-based environments. Wireless Communications and Mobile Computing, 2018, 6183732. https://doi.org/ 10.1155/2018/6183732 8. Dhanvijay, M. M., & Patil, S. C. (2019). Internet of things: A survey of enabling technologies in healthcare and its applications. Computer Networks, 153, 113–131. 9. Islam, M. M., Rahaman, A., & Islam, M. R. (2020). Development of smart healthcare monitoring system in iot environment. SN Computer Science, 1, 1–11. 10. Tamilselvi, V., Sribalaji, S., Vigneshwaran, P., Vinu, P., & GeethaRamani, J. (2020). Iot based health monitoring system. In 2020 6th international conference on advanced computing and communication systems (ICACCS) (pp. 386–389). IEEE. 11. Acharya, A. D., & Patil, S. N. (2020). Iot based health care monitoring kit. In 2020 fourth international conference on computing methodologies and communication (ICCMC) (pp. 363– 368). IEEE.

Application of Big Data Analytics and Internet of Medical Things (IoMT) in Healthcare with View of Explainable Artificial Intelligence: A Survey Anurag Sinha, Den Whilrex Garcia, Biresh Kumar, and Pallab Banerjee

1 Introduction Big data originally describe the volume, speed, and velocity of information that comes from different information creation time by medical services suppliers that contain data applicable to a patient’s consideration, including socioeconomics, analysis, operations, meds, crucial signs, inoculations, lab results, and radiology pictures. With the improvement of clinical information gathering, electronic wellbeing sources like sensor gadgets, streaming machines, and high throughput instruments are gathering more [1]. This medical care enormous information is utilized for different applications like the conclusion, drug discovery, accuracy of medication, expectation of infection, and so on. With the headway of society and the advancement of humanity, people see that prosperity is not only one of the targets social improvement pursues but also the fundamental state of propelling financial development. In this action, the clinical model has at this point changed from “biomedical mode” to “normal mental social clinical model,” moreover, altered therapy and a grouping of interventions have become the example of clinical new development [2]. As individuals’ requests for different clinical and medical care have been developing, the current general well-being administration

A. Sinha () Department of Information Technology, Amity University Jharkhand, Ranchi, India D. W. Garcia Department of Engineering, Lyceum of the Philippines University – Cavite Campus, Cavite, Philippines B. Kumar · P. Banerjee Department of Computer Science and Engineering, Amity University Jharkhand, Ranchi, India e-mail: [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_8

129

130

A. Sinha et al.

and its legitimacy have been enormously tested, and the deficiency of wellbeing assets and the insufficiency of clinical assets sharing have turned into the bottleneck of limiting its development. Every day, information is created by a scope of various applications, gadgets, and topographical exploration exercises for the motivations behind climate anticipating, climate expectation, catastrophe assessment, wrongdoing identification, also the heath business, to give a few examples. In the current situation, big data is linked to central developments, and Google, Facebook various efforts, including IBM; important information is taken from a very large collection of information. Open data for medical care is currently underway. Consideration by the patient, big data is quickly created for flexibility and different governance needs. Create creative ways to get a lot of new information and make the nature of medical work. A portion of the significant applications is general well-being, clinical tasks, innovative work, patient profile examination, proof-based medication, remote observing, and so on. The structures and capacity framework for medical services enormous information are represented underneath. Web of Things (IoT)—Open contraceptives offer distance vision in a practical domain; Experts are recruited to increase patient safety and likelihood and to consider the champion. As interviews with experts become more difficult and productive, responsibilities and patient satisfaction are taken into account in the same way. In addition, remote monitoring of the patient’s well-being makes it possible to reduce the duration of the crisis and prevent a recurrence. IoT additionally has a significant effect on decreasing medical care costs essentially and further developing therapy outcomes. Healthcare frameworks are by and large carefully changed by innovative upgrades in clinical data frameworks, electronic clinical records, wearable and keen gadgets, and handheld gadgets. This increment in large clinical information, close to the advancement of computational strategies in the field of medical care, has empowered specialists and experts to separate and envision huge clinical information in another range. In this advanced mechanical age, the data increments dramatically. Wearable gadgets constantly produce an immense measure of information that is at last known as large information in layman’s terms. Change is needed for the large information as a logical based procedure for legitimate administration, perception, and extricating the secret data inside the huge data [3]. Because the Internet of things innovation has clear benefits in the seeing, transmission furthermore, and use of data, its application in the field of clinical and medical services will benefit patients by obtaining the best clinical help, the base clinical expenses, the briefest treatment time, and the best wellbeing administrations. Through the Internet of Things, we can screen the entire course of creation, conveyance, hostile to fake and following of clinical hardware, oversee clinical data which incorporates ID, test ID, clinical records ID, etc., build administration frameworks that show restraint focused and in light of distant meeting and ceaseless observing of basically sick patients and develop a medical services the board stage, utilizing hardware which can sense, measure, catch, and communicate the data of the human body [4].

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

131

1.1 Problem Statement and Chapter Organization The problem with healthcare is managing the bunch of huge data and data which cannot be used; it becomes a burden for the warehouse. Similarly, IoT system in healthcare requires more precision decision support and analytics with artificial intelligence, which can only be possible through managing and mining useful information from the data warehouse. Thus, data fetching is very significant for analyzing big data for training the model, and getting insight from the raw data. In this particular chapter, a survey of big data with IoT implementation in healthcare is done. The first section of the chapter contains an introduction, background, and literature study. Furthermore, part of the chapter talks about various models of big data and IoMT that are useful for healthcare. Authors Contribution: The corresponding author of this has contributed the methodology and literature survey, and the second author of this chapter contributed the problem definition and background study of the chapter. The third author of this chapter has implemented the model study of big data and machine learning, and the fourth author of the chapter provided a relevant case study and helped in the chapter organization.

2 Background 2.1 Concept of IoT and IoMT in Healthcare The Internet of Things Concept The idea of “The Internet of things” was advanced in 1999. With a profound comprehension of the Internet of things, the implication of the Internet of things is more clear. The Internet of things can be just characterized as interfacing all things to the web through radio recurrence ID (RFID for short) and other data by detecting hardware to accomplish smart ID and the board. All in all, The Internet of things alludes to another innovation of joining a wide range of sensors and the current web. Web of Things upholds many info yield gadgets and sensors like cameras, receivers, consoles, speakers, shows, close to handle interchanges (NFC), Bluetooth, and so on. The fundamental part is the RFID framework. RFID can consequently distinguish the still or moving elements. The primary point is to screen and control objects through Internet [1–4]. The Medical Internet of Things—with the advancement of data innovation, the ascending of the electronic clinical, and the development of the Internet of Things innovation, the clinical Internet of Things has steadily incorporated into normal individuals’ life. The supposed clinical Internet of Things is a sort of innovation that installs remote sensors in clinical hardware, consolidates with the web, and incorporates with medical clinics, patients, and clinical gear to advance the new improvement of the present-day clinical model [5].

132

A. Sinha et al.

The clinical considerations manufacturing has not existed fast enough to adapt to the huge advances in data compared to various efforts. Similarly, the huge use of data in the field of clinical benefit is currently the most timely. For example, clinical benefits and huge biomedical data have not yet been combined to develop data on clinical considerations with nuclear pathology. Such a mixture can help to unravel different parts of the action or different pieces of impressive science. Therefore, biomolecular and clinical data sets should be discontinued in order to study the welfare of loners. One such source of clinical data for clinical benefit is the “case network” (IoT). In all honesty, IoT is another enormous player executed in various different endeavors, including clinical benefits. As of not long ago, the objects of normal utilization like vehicles, watches, refrigerators, and well-being checking gadgets did not generally deliver or deal with information and needed a web network. Nonetheless, outfitting such items with microprocessors and sensors that empower information assortment and transmission over web has opened new roads [6]. Advantages of IoT in clinical benefits utilizing the snare of IoT contraptions, an expert can measure and screen various limits from his/her clients in their singular regions, for example, home or office. Thus, with early intervention and treatment, it is unlikely that the patient will need a hospital or even a visit to a specialist to reduce the cost of medical reimbursement significantly. Some episodes of IoT devices used for clinical use combine health or well-being with highly qualified devices, biosensors, clinical devices to really see the underlying symptoms, and other types of devices or medical devices. Such IoT devices generate a lot of health-related data. If we reinforce this information and other considerations, such as EMR in medicine or PHR, we can best expect patient well-being and progress from the clinic to the local region. In fact, rogue data from the IoT have silenced the importance of a number of districts that provide good ratings and assumptions. To maximize, data from such devices can help ensure employee well-being by showing disease prevalence and finding ways to address windows [7].

2.2 Application of IoT in Healthcare Glucose control diabetes is a condition in which the body maintains its blood sugar levels for a long time. It is considered to be the most common contamination in an individual. The three main types of diabetes are generally straightforward and straightforward: type 1 diabetes, type 2 diabetes, and gestational diabetes. The analysis of IoT (Internet of Things) involves examining the data generated by interconnected devices and using it to extract insights that can inform decisionmaking and improve processes. However, the most illogical method used to diagnose diabetes is to test blood sugar. The analysis of IoT involves collecting and processing vast amounts of data generated by interconnected devices. This data can provide insights into patterns, trends, and anomalies that can inform decisionmaking and improve operational efficiencies. To perform IoT analysis, specialized

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

133

Fig. 1 Classification of IoT applications in healthcare

tools and techniques are used, such as machine learning, data mining, and predictive analytics. The insights gained from IoT analysis can be applied to a wide range of use cases, including supply chain management, predictive maintenance, and smart city planning (see Fig. 1) [8, 9]. Temperature monitoring: Human internal heat levels is a marker of homeostasis and are an important component of many symptomatic cycles. In addition, adjusting the internal heat level can be a warning sign in certain circumstances, such as injuries and sepsis. Long-term temperature control monitoring helps professionals draw conclusions about a patient’s medical condition for many infections. The traditional method of measuring temperature is to use a thermometer that is attached to the mouth, ear, or rectum. However, the patient’s low comfortability and large beats at the beginning of the disease are constantly a problem with these strategies. However, new developments in IoT-based innovation have offered different answers to this figure [10]. Blood pressure monitoring: An essential part of any indicator cycle is blood pressure (BP) prediction. A generally accepted strategy is to have at least one person perform a blood pressure test. However, the mix of IoT and other discovery innovations has changed the way BP has been noticed recently. For example, a displacement method that detects both systolic and diastolic pressure is recommended. Recorded data can be moved to the cloud. In addition, the effectiveness of this device has been tested on 60 people, and its accuracy has been confirmed [11].

134

A. Sinha et al.

Oxygen saturation, which measures heart rate, can be used as a key limitation in clinical trials, not as an assessment of oxygen uptake. The non-nasal mask eliminates the problems associated with conventional techniques. In rhythmic oximetry, which emerges from a mixture of IoT-based developments, the movement uses the potential in the realm of critical thinking. Decreased blood oxygen levels in the noninvasive tissue oximeter; it is recommended to assess heart rate and heart rate limitations. In addition, the recorded information can be sent by the employee using different communication enhancements, such as Zigbee or Wi-Fi. The practical intervention was determined by taking into account the recorded information. Medication management: Medication compliance is a matter of practice in practice. Failure to adhere to the medication plan can lead to the reverse surprise for patients. As you age, your mental health may decline. Medicines are found in the elderly to treat clinical conditions such as Alzheimer’s disease. Now it is difficult for them to follow the drugs of trained professionals strictly. Different surveys in the past have focused on IoT, wheelchair management using patient participation, and medication. Wheelchairs are an integral part of the limited availability of patients. It really gives them emotional support. In any case, wheelchairs are limited when power outages are the result of trauma. Later, the new assessment focused on the combination of these wheelchairs with the training and overall structure. The IoTbased infrastructure shows the expected results of achieving these goals [12]. In the field of clinical and medical services, significant utilization of the Internet of Things incorporate clinical gear and drug control, clinical data for the executives, telemedicine and portable clinical consideration, and individual well-being of the executives, which can be additionally clarified as the accompanying. Medical equipment and medication control: With the assistance of representation innovation of material administration, we can screen the entire course of creation, conveyance against fake and following clinical gear and medicine to shield public clinical well-being. In particular, the utilization of the Internet of Things in the checking and the executives of clinical hardware and prescription incorporate the accompanying viewpoints [7]. Constant real-time monitoring: In the entire course of prescription exploration, creation, dissemination, and use, RFID labels can be utilized to do all-adjust observing. Particularly, when meds are naturally pressed, the perusers introduced in the creation line can consequently distinguish the data of each medication and afterward communicate it to the information base. Available for use, the perusers can record all the data in the process whenever and do all-adjust checking. The drug quality can be ensured by observing medication conveyance and capacity climate. At the point when drug quality issues happen, we can follow back the deficient medication as per its name, classification, beginning, bunch, preparing, conveyance, stockpiling, deals, and other data. Medical refuse information management: With the participation of various clinics and transport organizations and the assistance of RFID innovation, a detectable clinical decline data framework can be set up, which can follow the clinical deny during the entire course of transport from emergency clinics to reject preparing

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

135

plants and keep away from unlawful removals of clinical reject. Right now, Japan has dispatched investigates around here and has made a few accomplishments [13]. Medical information management: The Internet of Things has expansive application possibilities in clinical data for executives. As of now, the requests of emergency clinics for clinical data the executives, for the most part, incorporate recognizable proof of patients or specialists, test ID of prescription, clinical gear or research facility synthetic compounds, clinical record ID of the condition of disease or signs, and so on Explicit applications include the accompanying a few viewpoints. Patient information management medical records and medical assessments: An electronic patient health profile, which includes medical records and medical care, will provide specific assistance in implementing treatment plans [13]. Over time, specialists and staff can continuously monitor the patient for essential symptoms and perform tests and treatments to avoid misuse or misdiagnosis [14]. Medical emergency management: Beneath a number of extraordinary conditions was harmed patients are excessive, patients’ relatives cannot be reached, or patients are in fundamental condition. With the help of strong and capable information storing and assessment procedures for RFID advancement, experts can quickly avow a patient’s character, which joins their name, age, blood grouping, telephone number, clinical history, relatives, and other organized information, furthermore, finish the enlistment approach to save the critical time emergency treatment [15, 16]. Blood information management: On the off chance that RFID innovation can be applied to blood the executives, it can viably stay away from the little limit disadvantage of standardized tags and acknowledge non-contact ID to diminish blood tainting, complete multitarget distinguishing proof, and increment the productivity of information assortment [17].

2.3 Big Data and Its Significance in Healthcare Important and practical information has recently appeared around the world. Basically, every survey produces big data for a variety of purposes, whether technical or not. Able to train the organization can store sensitive data. We need to know how it handles large amounts of data that are not normally programmable [18]. As a record of big data, medical considerations can be a major factor in the well-being of individuals. Protection against vulnerabilities is a multifaceted structure like diagnosis and treatment. The main components of the practical delivery system are wealth (specialists); build a financial environment that supports a thriving workplace (focus, overcoming drugs, and other motivations for seeking or treating). The prosperity specialists have a spot with various prosperity regions like dentistry, prescription, birthing help, nursing, mind examination, physiotherapy, and various others. Clinical consideration is required at a couple of levels depending upon the criticalness of the situation. Specialists serve as the chief place of conversation (for fundamental thought), extraordinary thought requiring gifted

136

A. Sinha et al.

specialists (helper thought), advanced clinical assessment and treatment (tertiary thought), and significantly remarkable logical or medical procedures (quaternary thought). Huge information initially represents the volume, speed, and assortment of information that comes from different information creation times by medical care suppliers that contain data applicable to a patient’s consideration, counting to socioeconomics, analysis, operations, drugs, fundamental signs, inoculations, lab results, and radiology pictures. With the improvement of clinical information gathering, electronic well-being sources like sensor gadgets, streaming machines, and high throughput instruments are collecting more. This medical services large information is utilized for different applications like finding, drug revelation, accuracy medication, forecast of illness, and so forth. Enormous information has been assuming a pivotal part in an assortment of conditions like medical care, logical exploration, industry, interpersonal interaction, and policy management [19]. Uses of medical care enormous information Enormous information applications present new freedoms to assorted new information and make inventive strategies to work on the nature of medical care. The significant applications portions are general well-being, clinical tasks, innovative work, patient profile investigation, proof-based medication, remote checking, and so forth. The structures and capacity framework for medical services huge information are represented beneath. Internet of Things (IoT): The included tools allow remote viewing in a practical area; it fulfills the possibility of keeping patients healthy and allows experts to consider the champion. Coordination with specialists becomes more complex and less efficient, increasing responsibility and patient satisfaction. In addition, seeing the well-being of the patient from a distance makes it possible to reduce the time spent in crisis and avoid recurrence. The IoT influences the reduction of the costs of medical services and the creation of outpatient treatment outcomes [20]. Advanced the study of disease transmission has characterized the study of disease transmission that utilizes computerized techniques from information assortment to investigate information. It improves customary epidemiological examinations, for example, case records, case reports, biological investigations, cross-sectional examinations, case-control considers, accomplice contemplates, randomized controlled preliminaries, and methodical audits and meta-examination. It additionally utilizes information sources that are initially gathered or potentially created for well-being and non-well-being-related purposes [21]. NoSQL data set is utilized to store a tremendous volume of information in a conveyed way. A NoSQL data set does not follow any social composition. NoSQL data sets can be grouped into four sorts, for example, key-esteem stores, section family information base stores, report stores, and diagram stores. The key-esteem stores the information dependent on key-esteem matches and is utilized for small applications. Besides, the segment family data set stores colossal information into lines as an assortment of sections. Large information might come from different sources like medical care, CCTV observation, social systems administration, machine-created information, and sensor information. The kind of information might be organized and unstructured. For

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

137

taking advantage of large information, there is a requirement for huge information engineering. Information in the request for 100 s of GB does not need any sort of engineering, yet when it goes past this, there comes a large information design. Interests in Big Data projects and have different wellsprings of huge information need enormous information design. Huge information engineering is intended to deal with the ingestion, handling, and investigation of information that is excessively enormous or complex for customary data set frameworks. Large information engineering is intended to deal with the accompanying kinds of work: • Batch handling of large information sources. • Real-time handling of large information. • Predictive examination and AI. Big data application in medical care The study of a large amount of information is a collection of resources and sizes ranging from terabytes to zettabytes. It uses unique strategies that combine fragmented and unstructured information. The vast scrutiny of medical information deals with the mixing and scrutiny of multiple heterogeneous complex information, for example, various information such as omitted biomedical information and HER (Electronic Health Record). It plays an important role among these EHRs and is widely accepted today in many countries. HER’s primary goal is to gain tremendous insight into the working process of wellbeing. EHR is an electronic (computerized) assortment of clinical data about an individual that is put away on a PC. An EHR incorporates data about a patient’s well-being history, like conclusions, meds, tests, hypersensitivities, inoculations, and treatment plans. EHR can be seen by all medical care suppliers who are dealing with a patient and can be utilized by them to assist with making suggestions about the patient’s consideration. EHR is likewise called EMR (Electronic Medical Record). Consistently, many terabytes of information are created and collected from different sources, e.g., web perusing, interpersonal organizations, versatile exchanges, web-based shopping, and numerous others. The big data worldview is completely ruined. In this way, the abundance of unstructured information gave us the idea of opening up new perspectives. AI for medical services big data Artificial intelligence is a subspace of automated thinking, whereby the term suggests the limit of IT systems to independently find deals with issues by seeing plans in informational collections. Artificial intelligence enables IT structures to see plans subject to existing computations and educational assortments and to cultivate agreeable course of action thoughts. Consequently, in AI, counterfeit data is delivered reliant upon experience. In AI, measurable and numerical techniques are utilized to gain from informational collections. There are two primary frameworks, specifically emblematic methodologies and subrepresentative methodologies. While emblematic frameworks are, for instance, propositional frameworks in which the information content, i.e., the instigated rules and the models, are expressly addressed, sub-representative frameworks are counterfeit neuronal organizations [22] (Fig. 2).

138

A. Sinha et al. Analytics Sensing Data

Descriptive

Diagnostic

Omics Data

Improved Outcomes Electronic Health Record

Data Warehouse

Public Health Record

Predictive

Smarter and Cost effective Decisions

Predictive

Clinical data

Fig. 2 Workflow of big data analytics in managing healthcare data

3 Literature Review This audit adds to the additional prosperity of the present by examining the actual transition of EHR’s logical gadgets in a superior data mode to form a significant use of massive data inspection with EHR, to organize special data [20]. In this chapter, they analyze key issues, data sources, techniques, progress similar to future topics in the field of huge evaluation of data from clinical benefit [21]. It conducts a DIY review that delivers a complete, improved, and hassle-free perspective on the various advancements used to encourage planned, successful, legitimate applications [22]. It describes a useful use of democratization of clinical violations for both patients and clinical benefit providers. In addition, we recognize the neglected evaluation process and potential examples for solving strange search problems. Formal pointto-point security is enforced, including BAKMP-IoMT security assertions using automatic validation of recognized Internet safety procedures and submission to make obvious adaptableness to dissimilar types of probable bother. The assessment of BAKMP-IoMT is coordinated with appropriate existing schemes recognizing that the proposed structure provides better security and utility and also requires a response and lower computational costs when highlighted in other schemes. BAKMP-IoMT entertainment is tuned to show the effect it has on display limits. In this chapter, reference [23] investigates the impact of massive data on clinical benefits and various disclosure violations in a natural Hadoop framework for therapy. Additionally, we explore a virtual plan for a massive assessment of data for clinical consideration that incorporates data gathered from various points, genomic information indexes, electronic records of prosperity, text/images, and emotionally supportive clinical decision networks. Break down the introduction of a brilliant energy consumption structure and similar proposed clinical considerations. We are also considering the introduction of the proposed EF-IDASC plot along with other related plans [22]. In this chapter, we plan to build a multicloud architecture to

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

139

create OpenStack-based steps for practical IoT related to the triple collection of frustrations (Tri-SFRS). To create Tri-SFRS, they used the multicolor drop design, low-end neighborhood test structure, integrate many of these efforts, including a support framework for convenient data storage and volume reduction. This chapter analyzes specific AI calculations applied to various clinical consideration data. In addition, there is the problem of planning and managing special data and its applications. The thesis aims to clarify the preconditions for the use of artificial intelligence computation and the processing and use of anomalous data presented from an alternative perspective. The convergence point of the thesis is big data evaluation and medical prescription [3]. Presenting and isolating the computational movement of data disclosure, we found valuable results for the logical examination of genomic sequences based on PC RNA models. Finally, useful occupations and related developments in the Internet of Clinical Objects were discussed. Another class of “Modified Security Prosperity Coaches” appears. These professionals will have secret sources and the ability to achieve and achieve thriving and thriving data. They will help clients avoid progress and diet-related disorders, work on mental abilities, achieve mental prosperity, and generally achieve an additionally created lifestyle. As people generally age, such positions will become dynamically important [16–21].

4 Big Data Analytics in Health Care Computer programming practices are the main difference between regular welfare assessments and ad hoc data welfare analyses. In the standard structure, the clinical considerations industry relied on a variety of organizations to evaluate large amounts of data. For many clinical benefits, financial sponsors are confident in disseminating information with significant results. The operating system can process data in a practical and standardized configuration. It is difficult for today’s clinical benefit industry to cope with the rapid development of large amounts of clinical benefit data. Surprising advances in clinical informatics have also been found in the field of bioinformatics, where genome sequencing results in other terabytes. Combinations of logical methods are available for clinical interpretation that can be used to consider the patient. At various stages of the initial and massive data types, attempts are made to consider computers in data preparation systems clinically. There is considerable interest in combining different data sources [24]. Perceptual Analysis in Health: Over the past 2 years, smart testing has been seen as one of the vast knowledge in the approaching business, yet its certified applications have relaxed much more than the business environment. The evaluation of exceptional data reinforces various methods, including text verification and combined media investigations. Nonetheless, perhaps the most horrific classification is the perception test, which combines realistic methods such as data mining and AI that divide current and real factors to anticipate what to expect.

140

A. Sinha et al.

Healthcare AI: The AI option is essentially equivalent to data mining capabilities [4], both of which do analyze data to capture plans. It is possible that by isolating the data subject for human understanding, such as in data mining applications, AI uses this data to advance the program design further. Computer intelligence perceives data models and changes the program accordingly in a short time. Electronic medical data: Tremendous information abundance is unavoidably utilized in the clinical assessment of EHR. Each open-mark individual has their own clinical records with subtleties that intensify their clinical history, end of touchiness, unplanned impacts, and lab test results. Patient enlistments are divided among both public and private districts, giving clinical advantage to suppliers through a protected data framework. These reports are editable, which implies that specialists can make changes after some time and add new fundamental clinical outcomes without utilizing archives or copying data [24].

4.1 Five Versus of Big Data Analytics in Health Care Four important tasks (shown in Fig. 3) related to large amounts of information: volume, speed, assortment, and truthfulness.

Terabytes Records Transaction Tables Files

Veracity

Trustworthiness Authenticity Origin Availability Accountability

4 V’s of Big Data

Volume

Structured Semi structured Unstructured All the Above

Variety Velocity

Fig. 3 Five Vs of big data in healthcare

Batch Near real time Real-time Streams

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

141

Scope: Big data is a term that encompasses huge amounts of aggregated data. There is no nice breaking point for this amount of data. This term is commonly used for gigantic data that needs to be monitored, maintained, and removed using standard indexes of information and data related to architecture. The amount of data provided by the current IT and clinical analysis system was established due to the low cost of data collection and planning structures and the need to isolate relevant information from the data to further enhance the customer business cycle, efficiency, and organizations [25]. Speed: Speed, which strives for fundamental legitimacy behind sensational data development, means how fast data collection is. Clinical investigation systems generate data at higher dynamic speeds. With regard to the amount and collection of accumulated coordinated or unstructured data, the rate of aging of such planning data requires a decision that depends on its effectiveness. Miscellaneous: Variety means unstructured text, practical images, sound, video and sensors. Harmonious data can be used to strengthen medical data (patient records); it must be maintained, especially as a tool. The traded data contains only 5–10% of the actual data. Unstructured or partial information is centralized; clinical records and special notes combines comfort information such as paper medications and X-rays. Truth: The truthfulness of the data is the degree to which the significance of the data is not surprising. The level of legitimacy and unwavering quality of data varies for different data sources. The consequences of gigantic data verification should be reasonable and free of charge, but for clinical benefit, independent AI calculations make a conclusion so as to be used with robotic equipment based on facts that may be unnecessary or misleading. The examination of a clinical trial depends on the removal of additional information from these data in order to extravagance patients and make the paramount pronouncement.

4.2 Benefit of Big Data on the Healthcare The power of large-scale data is the intention and can interfere with the most appropriate or precise patient diagnosis and accurate information used in large quantities of information. To this extent, extensive data analysis can effectively affect the effective structures of five-way organizations or “roads” [26]. Proper lifestyle: A healthy lifestyle refers to the patient’s transition to an unbeatable and better life. In real life, patients can administer themselves by creating the finest conclusion for themselves, considering information and making better choices, and making careful changes. Appropriate care: Dariiqaani ensures that patients seek treatment for the best available and that all providers have access to data comparisons and attention to avoid comparisons to looga -qabsiga and Hang -dhiska.

142

A. Sinha et al. Description Right Living

Right Care

Right Provider

Informed lifestyle choice that promote wellbeing and the active engagement of consumers in their own care Evidence-based care that is proven to deliver needed outcomes for each patient while ensuring safety Car provider (eg-nurse, physician) and setting that is most appropriate to deliver prescribed clinical impact

Right Value

Sustainable approaches that continuously enhance healthcare values by reducing cost at the same or better quality.

Right Innovation

Innovation to advance thee frontiers of medicine and boost R&D productivity in discovery, development, and safety

Ecosystem Feedback Loop

Fig. 4 Benefits of big data

Accurate Provider: Healthcare providers can use this to get an overview of their patients by combining data from a variety of sources, such as outpatient facilities, total population estimates, and financial data. Accurate creativity: This approach captures the development of new disease conditions, new treatments, and new health opportunities. Right Value: To chip away at the worth and assessment of prosperity-related parties, providers must provide careful and progressive thinking to their patients. Patients should get the most beneficial results recognized by their social security system [27] (Fig. 4).

4.3 Tools for Big Data in Healthcare For the healthcare industry, Hadoop-based applications need to digitally integrate print structure data, as healthcare information is critical to the printing environment. Much of this information is unconstitutional; therefore, this work is based on the thoughts and feelings of the patient. It is difficult to extract large amounts of information on operations and investigations. A variety of programming tools, known as Hadoop weather, can help manage these huge data ratios with practical benefits [28]. Cancer and Gene Therapy: We understand that human DNA contains three trillion bases. It is based on planning as much data as possible to fight against

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

143

malicious developments. The conditions that threaten developmental changes and their responses differ from the natural symptoms that cause some infections to be incurable. Medical care Intelligence: Hadoop development moreover maintains the clinical benefits od information applications used by crisis centers and protection organizations. Hadoop organic framework’s Pig, Hive, and MapReduce developments measure tremendous datasets related to medications, ailments, incidental effects, ends, and geographic regions. Medical clinic Network: Several clinical centers use the Hadoop climate’s NoSQL database to assemble and manage their gigantic proportions of nonstop data from arranged sources related to patient thought, reserves, and money, which helps them with perceiving high-risk patients while furthermore reducing regular employment. Avoiding and Detecting Fraud: In the early texts of massive data assessment, wealth-based security clusters used various methods to identify ways to recognize and prevent the development of blackmail and prevent fraud. Apache Hadoop: The name Hadoop is used to create a wide variety of things [23]. In 2002, he was named the only programmer who dared to help web browsers. Since then, it has become a natural program of programs and applications used to analyze all kinds of big data and categories. Hadoop might not be considered a solid business right now, but it is a way to deal with fundamental changes from the traditional model of social data collection. MapReduce: Apache Hadoop is often associated with MapReduce subscriptions. The MapReduce computational model is an incredible resource used in a variety of rich applications, surpassing most customers. Its central reflection is essential. HDFS: HDFS had to handle large amounts of data. He did not see HDFS as a fair documentation structure, no matter how he retained multiple clients. It is possible that program recognition will reduce the enormous requirements for a combination of upgrades and a recognized standard. Apache Hive: Hive is one of Hadoop’s most notable sites, where a repository is a place where you can take exams and quizzes using procedural language like SQL [29]. Apache Hive provides updated queries, and you can use individual data evaluations. Apache Pig: Apache Pig is an open-source platform that can be used to access large amounts of data faster. Pig is another alternative to the MapReduce programming gadget. Apache Avro: Avro is a PIPELINE program that allows you to exchange data between programs written in any language. This is the time it takes to redo the Flume data streams.

144

A. Sinha et al.

4.4 Dispute of Big Data in Healthcare Strategies for big data, on-board, and analysis are reliably implemented, particularly for continuous data transmission, retrieval, aggregation, analysis (using ML and predictability), and understanding game plans, which will assist in planning an unparalleled use of EMRs using clinical considerations. For example, the speed of EHR collection of administrative trials and asserted EHR programs in the United States clinical benefit region. It is almost complete [27]. Regardless of the fact, the opening up of many EHR objects expressed public strength, each with different clinical formulations, specific judgments, and useful capabilities, has caused interoperability difficulties and data exchange [30]. Limiting the deletion of a large amount of data is one of the major challenges, but different affiliates approve the collection of data in their own areas. It has many advantages, such as security control, access, and uptime. Regardless, the organization of groundwork can be overwhelming and difficult to figure out. Clearly, with declining costs and continued quality expansion, Cloud-based constraints using computer systems are the main choice chosen by the most practical beneficiaries. Clean data is a matter of accuracy, precision, and flexibility; it must be clean or disinfected to ensure suitability and failure of recovery. This cleaning frame can be manual or automatic and uses dimensions to increase accuracy and reliability. Newer and more accurate industry conferences are using AI technology to reduce time and costs and prevent the destruction of unsatisfactory data in one-off data projects. United patients gameplay produces a huge amount of data. It is very difficult to consider nighttime reflection on big data, especially when it comes to an optimal correlation of data with providers of clinical considerations. The need to organize each of the relevant clinical information emerged along with the ultimate purpose of the cases, billing objectives, and clinical evaluation. Preliminary imaging processing studies have identified diverse real components that may induce adjusted data quality and misinterpretation of existing clinical records [31]. Clinical imaging continues to suffer from technical limitations that include different types of excitement and trinket. Inadequate treatment of clinical images can cause image changes. Metadata have a successful data organization plan; having accurate and compelling metadata is essential for all the information you have. Metadata contains information about past usage (who, why, how, and when) for professionals and data professionals. It will allow experts to copy previous questions and help with more cognitive analysis and accurate comparison tests. Portraiture: A perfect and attractive description of the data with traces, heat guides, and histograms to display separation numbers and proper stamping of the information to reduce the potential for mixing can be made more complex for us to ingest the information and use it properly. Different models adopt bar contours, pie charts, and diffuse plots with their own specific ways of coping with data transmission.

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

145

5 Internet of Medical Things in Health Care 5.1 Challenges for IoMT Driving IoT stages should give straightforward, amazing application admittance to IoT gadgets and information to assist designers with quickly making examination applications, perception run sheets, and IoMT applications. Coming up next are five key abilities that driving stages should empower: Simple network: A decent IoT stage makes it simple to interface gadgets and perform gadget the board capacities, scaled through cloud-based administrations, and apply investigation to acquire knowledge and accomplish the hierarchical transformation. Easy gadget the board: A smart way to deal with gadgets the executives empower further developed resource accessibility, expanded throughput, limited spontaneous blackouts, and decreased upkeep costs. Information ingestion: Intelligently change and store IoT information. APIs connect the split between the information and the cloud, making it simple to pull in the information that is required. Information is ingested from different information sources and stages, and then, at that point, the fundamental qualities are separated utilizing rich investigation. Informative investigation: Familiarize yourself (acquire, obtain) with current techniques from transcendental IoT. Respond to the same situation using continuous testing to verify the current situation. Conditions both systematic and unstructured factors influence the learning process from the choices made as circumstances change. A natural dashboard makes everything easy to understand. Reduced danger: Act on notices and disconnect occurrences produced anyplace in the organization climate from a solitary control center [31] (Fig. 5).

Fig. 5 IoMT challenges

146

A. Sinha et al.

5.2 Impact of IoT in Healthcare The IoT is relied upon to engage a collection of clinical consideration organizations where each help gives a lot of clinical consideration game plans. Concerning clinical consideration, there is no standard importance of IoT organizations. Be that as it might, there may be a couple of situations where help cannot be evenhandedly isolated from a particular plan or application. This chapter suggests that help is by sure means customary in nature and might conceivably be a construction block for a lot of game plans or applications. Moreover, it should be seen that general ser obscenities and shows required for IoT frameworks may require slight adjustments for their authentic working in clinical benefits circumstances. These join notice organizations, resource sharing organizations, internet services, crossaccessibility shows for heterogeneous devices, and association shows for the critical organization. The basic, speedy, secure, and low-power disclosure of contraptions and organizations can be added to this summary. Regardless, a discussion on such summarized IoT organizations is past the degree of this review. The fascinated examination has implied the composition for a more careful under-staying of this point. The going with subsections consolidates various kinds of IoT clinical consideration organizations [32] (Fig. 6).

Fig. 6 IoT in healthcare services

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

147

6 Machine Learning for Big Data Machine intelligence has long been used to develop drug display devices. It is accepted that no more work can be done to properly use the daily looga routine. As part of the data acquisition, ML is found in the data process for data collection and grouping on the device. In this case, the central problem arises from individual data changes and proportional distribution of data. We use ML to determine the accuracy of this data to identify different patterns in our data and develop treatment tools for future data research. Developing an effective tool will help you think more about health. We used data that included both the early stages and the minority of people with disabilities. We use libraries like Corner to enhance personal data resources. The main factors to consider in this survey are age, male and female location, and social events in the blood. The purpose of the rule is to synchronize the different data and show the program [29]. Health insurance covers photos, coordinates educational records, such as genetic information and EP information. In practice, ML cadres attempt to score points or cut out the likelihood of a condition outcome. As subunclear drugs and prescriptions develop, AI is being developed into specific drugs, frontal cortex; nanomedicine can be used. Human intelligence systems are designed to diagnose patients using specific stimulus therapies using AI technology. Diagnostic or computational approaches diabetes heart disease. It has been developed to detect skin diseases. Delhi shows a bright map over 3 months to test the air quality index (AQI). This assessment was conducted by India Today, a data channel news media organization, and shared obscene information and knowledge from the CPCB. The CPCB is a member of the Central Committee for Environmental Pollution Control; it is the official liaison office of the Ministry of Environment and Forestry. Therefore, this can be undetectable, no matter how unpredictable, with unpredictable data effects. Unavailable information. With reference to Dataset, we have compiled this data collection through the google architecture strategy we have had with our classmates and peers. That is the explanation; most survey data have individual patient data from packs between the ages of 16 and 30. After obtaining complete answers, further treatment interventions include BMI calculations in height and individual weight data, replacing parts such as medical history, symptoms screening, etc., in the Boolean game plan, near the time of the social event of individuals entering a 2-year gathering plus one session. Data collection planning contributed to the right, unobtrusive presentation and implementation of ML data tools [33] (Figs. 7 and 8). The above diagram illustrates the analysis of respect for gender and the analysis of male and female expressions. The red dots represent women, and the blue dots represent men. In these signals, we can understand the opposite of the two axes [34] (Fig. 9). The histogram above gives you an overview of the life of internal organs. Compared to the main data, the data shows that members are generally between 18 and 21 years old. It is based on the fact that most of the young members have

148

A. Sinha et al.

Fig. 7 Flow of algorithm and dataset sample

acquired a high level of education since the study. This helps us to anticipate what to expect in general, right down to the most common types of diseases and clinical dates in contact with humans. It is the most famous blood of all Indians. He goes into great detail about the accuracy of this information [35]. The heat map provides information about the types of information collected from the exam. The dark green shade provides the standard information type for information. Therefore, light shades indicate inferior information that must be processed before the visual inspection. This test was performed to determine the effect, without further study. Think of several numbers backwards. Complete information [36] (Fig. 10).

7 Proposed Big Data Analytics and IoMT Architecture for Health Informatics 7.1 Application of an IoMT Structural Design and Stand to Facilitate Invasive Healthcare This section identifies the proposed IoMT functionality, taking into account the above issues. Figure 3 shows the main elements of the cloud and the IoMT series in the image diagram, showing the beginning of the scene. It is most likely that the data will be collected by various sensors or taken from patient review documents and moved from the fog to the clouds at the entrance. Patient cells are thought of as an input to detect factors in the patient’s environment. At the same time, workers are sent to form centers or offices that receive a portion of the health information system in the area. It integrates a cloud that is ready to retrieve data and analyze experiences. Information based on AI programming and web analytics [37].

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

149

Fig. 8 Data relation with respect to gender using pairplot function of seaborn library Visualizing Important Features ca oldpeak thalach cp

Features

thal age trestbps chol slope exang sex restecg fbs 0.00

0.02

0.04

0.06

0.08

Feature Importance Score

Fig. 9 Histogram visualization

0.10

0.12

0.14

150

A. Sinha et al.

Fig. 10 Heat map visualization

In an overview of the program, the stage is shown in Fig. 4 with 50% of the windows shown on the left, with a compromised layer; presented with data integration layer and retrieval and representation layer. Device Interaction Layer: This layer integrates a variety of sensors and other sources of health information (e.g., other abnormal EHRs or SIS). Three different types of data sources and developments are considered. And (3) a component of the data source for the IoMT phase; It is part of the data resources for the IoMT level, similar to the IoMT levels open to SIS organizations for health benefits and other health care inputs. Data Integration Layer: The data integration layer is the data, support essential for management and access. The CSE obtained from OneM2M supports a number of organizations that collect data from OneM2M IoMT disputes. Adapters, on the other hand, promise to send data collected from various sources. Both are data coordinators; the OpenEHR and OpenEHR problem is reinforced by a combination of data. Data Recovery and Data Viewing Layer: This coating offer apparatus for retrieving appointments and displaying data. The data representation element enables data analysis by visual interaction using 3D OLAP. In line with these channels, experts can request the volume of data collected by IoMT organizations and various troublefree structures. IoMT Device (DNA): Because IoMT gadgets follow the specifications of OneM2M, we offer an enhancement system written in a worldview article that

Data Marts/OLPA Server

Data Mining Tools

FHIR API

Data Wrangling

OPenEHR To FHIR

OpenEHR Storage Data Warehouse

OpenEHR Repository

Data Coordinator

In cloud or External

Data Visualization (OLPA)

Cloud Services

Applications (i.e Clinical Decision Support)

Domain Experts (Physicians/Developers)

151

IN-CSE

IoMT Gateway (MN) IoMT Devices (AND)

IoMT Fog Services (MN) IoMT Platforms

IoMT Edge/Fog Services Services

Devices Integration Layer

Data Integration Layer

Knowledge Extraction/Data Visualization Data Layer Consumers

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

Fig. 11 Architecture of framework

works in the AE development cycle and required protocols (BP). Figure 6 shows its major classes. The stage naturally prepares information for OLAP use by anonymity in records, thereby allowing patient protection. The inclusion of an appropriate Hadoop document framework improves the use of information exploitation, information retrieval techniques, and preparation of AI models, relying on the initial manipulation of anonymous records that occurs due to the use of calculations of map/reduce archive of medical services. An advancement structure has been added made for IoMT gadgets, penetration, and fog, for the incorporation of new-fangled sensors on stage, the expansion of AE functions in the path, and the implementation of new restriction connections. BP’s advancement of focused information commodity is empowered healthcare professionals rather than separate gadgets. Finally, the presentation tests gave excellent results and approved the stage in many of the scenarios under consideration lowest rest. Regarding REST BPs, dormancy has been shown to be satisfactory and considering the beneficial benefits of MQTT [35] (Fig. 11).

152

A. Sinha et al.

7.2 Architectural Framework for Big Data in Healthcare The theoretical framework for the explicit validation of data into practical considerations is similar to the Rich Data Project or Standard Assessment. The key skills depend on how you manage them. A classic heritage survey project can be carried out with a business intelligence tool presented in an autonomous structure, such as a workplace or a PC. Because a lot of information has such important meaning, care is taken at different data points. The positive result of the judgment has been significant for a long time. The big news is that medical thinkers are using their huge data warehouses to understand decision-making about the wealth of education better. In addition, open source steps such as Hadoop/MapReduce work in the cloud and use big data analytics for practical benefits. Although the estimates and models are virtually the same, the user interface of traditional test equipment and the features used for big data are absolutely unique. The advanced analytics gadgets are easy to use and straightforward. The practical benefits include the consistency of big data (e.g., electronic records of wealth, strong decision-making bodies, CPOE) and external resources (government resources, research institutes, pharmacies, defense offices, and HMO), and different fields (such as objections from different practical beneficiaries) including different inheritance and applications resources and data types [37]. Web and electronic media data: Click stream and collaboration data from Facebook, Twitter, LinkedIn, online diaries, etc. It can similarly join prosperity plan locales, mobile phone applications, etc. • Machine to machine data: readings from far-off sensors, meters, and other fundamental sign devices. • Big trade data: clinical benefits claims and other charging records are continuously available in semi-coordinated and unstructured setups. • Biometric data: fingerprints, inherited characteristics, handwriting, retinal scopes, x-pillar and other clinical pictures, circulatory strain, heartbeat and heartbeat oximetry readings, and other similar sorts of data. • Human-created data: unstructured and semi-coordinated data like EMRs, specialist’ notes, email, and paper reports. In this next part of the calculated structure, a few choices are made with respect to the information input approach, conveyed configuration, device determination, and investigation. Hadoop Distributed File System (HDFS): HDFS can encrypt the Hadoop package. It divides the data into lower-level sections and distributes them to different workers/departments. Finally, the models incorporate four routines that include big data validation on the right side. These questions reports integrates OLAP and data mining. Insight is a general theme in small business and productivity events. When implementing efforts, MapReduce follows each employee/hub’s plan. Computer science estimates, taking advantage of areas such as applied mathematics and financial problems; integrity of techniques and subjects; control; it has changed based on research and practical considerations. PIG and PIG Latin (Pig and

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

Big Data Sources

Transformed Data

Big Data Platforms & Tools • Hadoop

Middleware

• Pig Raw Data

Extract Transform Load

Transformed Data

• Multiple Formats

• Multiple Locations • Multiple Applications

Big Data Analytics Applications Queries

• MapReduce

• Internal • External

153

• Hive • Jaql

Big Data Analytics Reports

• Zookeeper Data Warehouse

• Hbase • Cassandra

OLAP

• Oozie • Avro Traditional Format CSV, Tables

• Mahout

Data Mining

• Others

Fig. 12 Architecture of framework for Big Data

PigLatin) web programming languages are intended to hold a wide range of information (integration/structure, etc.); it includes a real language called PigLatin and a modification of the working time in PigLatin. Running the Code The basic step for evaluating very large data is open source circled data, the first step in Hadoop (Apache step), the first step in normalization, such as collecting search records on the Web. This is especially true for “NoSQL” level enhancements, especially with data level upgrades. Then in step 3, the methods of the framework are identified and implemented. The expression of ideas is classified as a movement of ideas. (They do not realize that this is possible through quantitative approaches. They are probably designed to help coordinate big data analysis metrics). In the meantime, understand the neighborhood and the variables or factors. The data sources illustrated in Fig. 12 are better recognized. Collect information; demonstrated and modified it for examination. The most important step now is the step/instrument evaluation and decision making. There are two predetermined decisions: AWS Hadoop, Cloudera, and IBM Big Insights. The resulting step is to apply a wide variety of data validation techniques to the data. This communication simply transforms strategies to extend to a wider range of learning styles. Think about how you can assess through meaningful movement; Get insights from big data analytics Decisions can be made based on information. In step 4, the models and their expressions are familiar with the concepts and the efforts to support them. The execution is a system of circles of analysis formed at each stage to limit the risk of frustration. The content area illustrates two great data auditing applications found in practical benefits. We pull straightforward open source content from a variety of sources, including dealer objections. There is a weakness in this emerging discipline for assessing weaknesses. These models come from support sources. Coincidentally, they demonstrate the potential for large-scale data assessment in practical terms [38–40].

154

A. Sinha et al.

8 Challenges with the Implementation of Big Data and IoMT in Healthcare: Use of Big Data in IoMT Requires Understandable and Cognitively Stable Systems for the Future The approachable proposed model and explanation of what we have observed throughout the chapter, all require the system to be more interpretable, expert, and cognitive, so in this section, we are discussing how the combination of explainable AI with big data and IoMT in healthcare can solve this issue [41].

8.1 The Four Aspects of AI Diagnostics [42] A patient-centric system to explainable AI The development of machine learning/deep learning styles has great potential for drug discovery and treatment programs. There is a huge influx of environment around these developments, but there are some issues with the requests. Many algorithms in AI solutions are FDA-approved, and research is growing rapidly elsewhere. Explanation and effective competition Patients can be guided by an explanation of their AI-generated diagnoses with different principles and interests. Defining the need for an explanation usually means that the symptoms are symptomatic; the signs and indications must be the best possible scientific explanation. Some definitions would require a statement for the effect that the diagnosis came from the machine based on health data. Four faces of the effective competitiveness of AI diagnostics • The use of personal health data for AI diagnosis: First, competes with the use of personal health data in AI driving tests. • The AI diagnostic model uses personal health data from different sources. This personal health information is not only sensitive but also sensitive. Out of date, kill one side; this can also be wrong and perfect. Therefore, AI diagnostics cannot only lead to the unwanted use of sensitive information but also to the use of inaccurate personal information, which can lead to the use of inaccurate personal health data. Future AI systems can also use unhealthy data. Social conditions: this raises additional concerns regarding data resources, quality, and reliability. • Potential bias in the diagnosis of AI: Second, you need to be able to compete with potential biases in diagnosing AI. Recovery of AI data may be due to bias in training data or prior classification of training data. Bias has the right to protect individuals from discrimination; Therefore, AI should be able to compete with secularism.

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

155

• Third, you have to be able to compete with the diagnostic capabilities of AI: the AI model “locked” for diagnostics will have a capacity when it uses the same number of patients as the original learning package. Although there are many models of AI superior to humans, they are not completely accurate. They will inevitably produce over/under-diagnosis errors and omissions for an individual, which could possibly be dangerous. The capabilities of an AI model: for example, the number of true and false positives and the number of true and false negatives. • Sharing with the Union of Diagnostic Workers: fourth, a person should be able to compete in diagnostics between healthcare professionals. Either level of AI incorporation into the diagnostic process may vary depending on the distribution of diagnostic personnel between the system and the HCP. Direct integration into the diagnostic process for initial screening before diagnosis; Diagnosis required by the health professional; It can be used as second sight or in other words. The division and analysis of diagnostic staff between AI and professionals can hinder or improve the quality of diagnostic processes and improve the personal diagnostic decisions of healthcare professionals [42].

8.2 AI Effectiveness in IoMT Decision Making using XAI [43] “Algorithmic Efficiency of Explainable AI for Decision Making” refers to the importance of designing machine learning models that can not only produce accurate results but also do so in a way that is efficient, scalable, and interpretable. In other words, the focus is on creating machine learning models that can provide accurate predictions while also being transparent and understandable to humans, especially in decision-making contexts where transparency and interpretability are critical. Explainable AI models are gaining traction in many areas, including healthcare, finance, and autonomous vehicles, where the outcomes of machine learning models need to be transparent and explainable. However, explainable AI models often come at the cost of computational complexity and decreased performance. Therefore, balancing algorithmic efficiency and explainability is a crucial consideration in designing machine learning models for decision-making applications. Ultimately, the goal is to design AI systems that not only produce accurate results but also help humans make informed decisions by providing insights that are easy to understand and act upon [44].

8.3 Decision Making of XAI The decision-making process in XAI (Explainable AI) involves interpreting the output of the machine learning model to understand how it arrived at its predictions or decisions. XAI models are designed to provide interpretable output that can be

156

A. Sinha et al.

easily understood by humans, especially in complex decision-making contexts such as healthcare, finance, and autonomous systems [45]. The decision-making process in XAI involves multiple steps, including data preparation, model training, model evaluation, and model interpretation. The interpretation step is critical in XAI because it provides insights into how the model arrived at its decision or prediction. Interpretability is achieved by designing XAI models that are transparent, explainable, and easily understandable by humans. To make effective decisions with XAI, it is important to understand how the model arrived at its decision and whether it was influenced by factors such as bias, data quality, or model complexity. By understanding these factors, decision-makers can improve the performance of XAI models and make more informed decisions. Ultimately, the goal of XAI is to create models that produce accurate results while also being transparent, interpretable, and trustworthy [46].

9 Explainable AI in Healthcare [47] Explainable AI (XAI) is an emerging field in artificial intelligence that aims to make machine learning models more transparent, interpretable, and explainable to humans. In healthcare, XAI can help clinicians make more informed decisions by providing them with insights into how machine learning models arrived at their predictions or recommendations. This can improve patient outcomes and help healthcare providers optimize their workflows. XAI can be used in several healthcare applications, such as medical imaging, diagnosis, drug discovery, and treatment planning. For example, in medical imaging, XAI models can help radiologists interpret complex images by highlighting areas of interest and providing explanations for the model’s predictions. In diagnosis, XAI models can help clinicians make more accurate and timely diagnoses by providing them with insights into the factors that influenced the model’s decision. In drug discovery, XAI models can help researchers identify new drug candidates and predict their efficacy based on large and complex datasets. One of the main benefits of XAI in healthcare is its potential to reduce bias and improve fairness. By providing transparency and interpretability, XAI models can help identify and mitigate sources of bias in the data or model architecture. This is particularly important in healthcare, where decisions can have significant consequences for patients. Despite its potential benefits, XAI also poses challenges in healthcare, such as data privacy, model complexity, and regulatory compliance. Therefore, it is important to develop ethical guidelines and best practices for the responsible use of XAI in healthcare. In summary, Explainable AI has the potential to revolutionize healthcare by improving clinical decision-making, reducing bias, and enhancing patient outcomes. However, it is important to address the challenges and ethical considerations associated with XAI to ensure that it is deployed in a responsible and safe manner.

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

157

Fig. 13 Integrated XAI Framework for Big Data and IoMT in healthcare [47]

9.1 Interpretable Machine Learning in Healthcare [43, 46] Defined machine learning refers to machine learning models that can explain some of the predictions made. It is not enough to simply support traditional machine learning metrics, such as AUC, in many areas where there is a need to rely on the predictions of machine learning systems. Although machine learning techniques have been used for decades, the expansion of these technologies into areas such as healthcare has placed more emphasis on machine learning systems. Health care practitioners and other decision-makers prioritize the implementation and enforcement of standards. As machine learning applications become more and more integrated into different components of patient care, estimates are needed. In addition to therapeutic areas, machine learning solutions are used to help with surgery and costs. Decisions based on machine learning projections are based on diagnosis; it can also provide information on clinical procedures and patient risk assessment. Doctors and others want to know the reasons behind the estimates of such referral decisions. In this lesson, we will give a complete overview of the differences in explanations in machine learning. In this lesson, you will learn about different aspects of machine learning explanation. The explanation is not only for machine learning but also for input data and other aspects of machine learning, such as model parameters and algorithms. The type of explanation is highly dependent on the system user. For example, capacity (perception), inexperienced expert (domain knowledge), explanation in-depth, the future of machine learning and healthcare covers some of the current trends. It explores some of the emerging areas in this field by exploring and identifying the areas that benefit the most from learning from the application of the field of health education [46]. Wanting to improve machine learning models does not always reflect the real desire for the job at hand. Defining models provides information that evaluates our decisions and does not serve a single purpose. Definitions can be difficult to define in many ways. First, explore definitions that can be used to compare frameworks and design interpretations in general. Practice machine learning techniques to improve

158

A. Sinha et al.

precision-based matrices. In many cases, the function of a goal does not accurately capture the true costs of modeling decisions. Costs related to ethics or justice are difficult to define in a mission, and researchers may not be able to prioritize these costs [46]. If the measure is not sufficient, a definition is needed. The definition is the study of an exact form; all of this allows us to assess all of this in the context of the realworld problem that we are trying to solve in order to understand the justice behind the model’s other information and its decisions. Trust definitions can define human beings as a requirement for believing in role models. However, trust, like the definition, is difficult to define. One way to define trust is to appear to match the standard spread in the real world. We may be more comfortable with a simpler model, especially for high-stakes betting decisions (like money and drugs), but how precise is that comfort model? This does not necessarily reflect efficiency. In terms of the trust, we care how bad the models are. In particular, we have a strong desire not to imitate human error. Additionally, we seek to believe that the model will continue to work when the deployment differs from the training environment. Transferability: Humans generally have superior capabilities to machine learning modes. When the test environment changes from the training environment, the models help to understand how the models can be trained. In some cases, this change is a natural consequence of the data itself or the result of a change in the environment because of the propagation of the model [46]. Information: In many cases, the purpose of optimizing the model is to serve as weak oversight for the real purpose of providing useful information to human decision-makers. The form uses the direction of the search object. Examples include intermediate symptoms learned from convolution networks and parameter intensities in the form of linear regression. You need to make sure you fully understand the learning model (ML). The high definition of the design makes it easier for users to understand and explain future forecasts. In addition, meaningful ML forms allow healthcare professionals to make informed decisions about improving the quality of healthcare services. In general, the first group distinguishes two groups that estimate population levels (overall interpretation) in terms of interpretation. In other words, we can group together groups of methods designed to interpret the predictions of a specific model, such as a neural network. The model for IoT (Internet of Things) involves designing and deploying interconnected devices that collect and transmit data to a centralized system. This system then processes and analyzes the data using various tools and techniques, such as machine learning and data analytics. The insights gained from the analysis can be used to optimize processes, improve operational efficiencies, and inform decision-making. However, designing and deploying an effective IoT model requires careful consideration of factors such as security, scalability, and interoperability [42, 43] (Fig. 14). Eight Comparative Study and Research Gap Identification (Table 1) In paper 1, the author has done a survey for big data management; however, some models could have been introduced. In paper 2, authors had given a theoretical

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

159

Trust Cansality Informativeness Transferability Fairness

Validation Performance x1 . . xn

Model-Agnostic Model-specific

Data

Model Global

P P4 P5

P2

P1 P3

P6

P4 P5

Humans Local

P

P2

P1 P3

Interpretability Methods

P6

IF AGE > 50 AND BP > 130 AND FPGL < 6.1 THEN TRUE

P P2

P1 P3

P4 P5

P6

Fig. 14 Interpretable ML meaning and machine learning models for calculation in healthcare based on their interpretability characteristics

Table 1 Papers for study

Paper title “A comprehensive analysis of healthcare big data management, analytics and scientific programming”

Work done Survey of big data management

“Big data analytics in healthcare: Promise and potential”

Theoretical framework for big data

“IoMT platform for pervasive healthcare data aggregation, processing, and sharing based on OneM2M and OpenEHR”

Fast healthcare interoperability resource (FHIR) application programming interfaces (APIs).

“A scalable multicloud storage architecture for cloud-supported medical internet of things”

Scalable multicloud storage architecture

“Big data, big knowledge: Big data for personalized healthcare” IEEE Internet of Things Journal, Vol. 5, No. 5, October 2018

Big data technologies for in silico medicine

Big data in healthcare: Management, analysis and future prospects

IoMT-targeted microarchitecture research

Integration of biomedical and healthcare data

160

A. Sinha et al.

framework, but some practical frame should have been given. In paper 3 FFHIR model was made; however, system security risk is still a big concern. In paper 4, cloud architecture was shown; however, it needs to be collected with IoT node. In paper 5 authors have given research architecture, but security concern is missing and should be implemented. In paper 6, the author has given integrated framework, but big data security concerns should have been highlighted.

10 Conclusion In conclusion, the integration of Big Data Analytics (BDA) and the Internet of Medical Things (IoMT) in healthcare has brought significant benefits in improving patient outcomes, reducing healthcare costs, and enhancing personalized medicine. However, the use of BDA and IoMT in healthcare also poses challenges such as data privacy, security, and regulatory compliance. Explainable Artificial Intelligence (XAI) has emerged as a solution to address these challenges by providing transparency, interpretability, and fairness to machine learning models. XAI has the potential to revolutionize healthcare by providing clinicians with insights into how models arrived at their predictions or recommendations, thus enhancing decision-making. This survey has presented an overview of the latest research and development in the integration of BDA, IoMT, and XAI in healthcare. It has highlighted the various applications of BDA, IoMT, and XAI in healthcare, such as medical imaging, drug discovery, diagnosis, and treatment planning. The survey has also discussed the potential benefits of XAI in healthcare and the need for ethical guidelines and best practices for the responsible use of BDA, IoMT, and XAI in healthcare. In summary, the integration of BDA, IoMT, and XAI in healthcare has the potential to revolutionize the healthcare industry by improving patient outcomes and reducing healthcare costs. However, it is important to address the challenges and ethical considerations associated with their use to ensure patient safety and privacy. The development of XAI has provided a solution to some of these challenges, and future research should focus on the responsible use of BDA, IoMT, and XAI in healthcare to achieve the full potential of these technologies.

References 1. Yang, Y., Zheng, X., Guo, W., Liu, X., & Chang, V. (2019). Privacy-preserving smart IoT-based healthcare big data storage and self-adaptive access control system. Information Sciences, 479, 567–592. https://doi.org/10.1016/j.ins.2018.02.005 2. Zhen, M. (2012). Application of big data in public transportation. In 2021 international conference on information technology, education and development (ICITED 2021) (pp. 22– 28). https://doi.org/10.25236/icited.2021.019.

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

161

3. Rubí, J. N. S., & Gondim, P. R. L. (2019). IoMT platform for pervasive healthcare data aggregation, processing, and sharing based on oneM2M and openEHR. Sensors (Switzerland), 19(19), 1–25. https://doi.org/10.3390/s19194283 4. Dash, S., Shakyawar, S. K., Sharma, M., & Kaushik, S. (2019). Big data in healthcare: management, analysis and future prospects. Journal of Big Data, 6(1), 1–25. https://doi.org/ 10.1186/s40537-019-0217-0 5. Liu, J., Ma, J., Li, J., Huang, M., Sadiq, N., & Ai, Y. (2020). Robust watermarking algorithm for medical volume data in internet of medical things. IEEE Access, 8, 93939–93961. https:// doi.org/10.1109/ACCESS.2020.2995015 6. Savage, N. (2012). Digging for drug facts. Communications of the ACM, 55(10), 11–13. https:/ /doi.org/10.1145/2347736.2347741 7. Nazir, S., Ali, Y., Ullah, N., & García-Magariño, I. (2019). Internet of things for healthcare using effects of mobile computing: A systematic literature review. Wireless Communications and Mobile Computing, 2019. https://doi.org/10.1155/2019/5931315 8. Kjeldskov, J., & Skov, M. B. (2007). Exploring context-awareness for ubiquitous computing in the healthcare domain. Personal and Ubiquitous Computing, 11(7), 549–562. https://doi.org/ 10.1007/s00779-006-0112-5 9. Sadiku, M. N. O., Eze, K. G., & Musa, S. M. (2018). Wireless sensor networks for healthcare. Journal of Scientific and Engineering Research, 5(7), 210–213. Available online www.jsaer.com. (Wireless sensor networks for healthcare. (2018, September), pp. 3–7). 10. Lukowicz, P., Kirstein, T., & Tröster, G. (2004). Wearable systems for health care applications. Methods of Information in Medicine, 43(3), 232–238. https://doi.org/10.1055/s-0038-1633863 11. Mary, A. V., & Jerine, S. (2020). Wireless body area network transmissions for iot-based healthcare network: A review. IOP Conference Series: Materials Science and Engineering, 983(1), 012017. https://doi.org/10.1088/1757-899X/983/1/012017 12. Singh, S. (2021). A clustering-based optimized stable election protocol in wireless sensor networks. In Applications in ubiquitous computing (pp. 157–176). Springer. 13. Pradhan, B., Bhattacharyya, S., & Pal, K. (2021). IoT-based applications in healthcare devices. Journal of Healthcare Engineering, 2021. https://doi.org/10.1155/2021/6632599 14. Khatoon, N., Roy, S., & Pranav, P. (2020). A survey on applications of internet of things in healthcare. Intelligent Systems Reference Library, 180(June), 89–106. https://doi.org/10.1007/ 978-3-030-39119-5_6 15. Bensaleh, M. S., Saida, R., Kacem, Y. H., & Abid, M. (2020). Wireless sensor network design methodologies: A survey. Journal of Sensors, 2020. https://doi.org/10.1155/2020/9592836 16. Portocarrero, J. M. T., et al. (2014). Autonomic wireless sensor networks: A systematic literature review. Journal of Sensors, 2014. https://doi.org/10.1155/2014/782789 17. Islam, S. M. R., Kwak, D., Kabir, M. H., Hossain, M., & Kwak, K. S. (2015). The internet of things for health care: A comprehensive survey. IEEE Access, 3, 678–708. https://doi.org/ 10.1109/ACCESS.2015.2437951 18. Zhang, C., Ma, R., Sun, S., Li, Y., Wang, Y., & Yan, Z. (2019). Optimizing the electronic health records through big data analytics: A knowledge-based view. IEEE Access, 7, 136223–136231. https://doi.org/10.1109/ACCESS.2019.2939158 19. Hu, F., Xie, D., & Shen, S. (2013). On the application of the internet of things in the field of medical and health care. In Proceedings – 2013 IEEE international conference on green computing and communications and IEEE internet of things and IEEE cyber, physical and social computing, GreenCom-iThings-CPSCom 2013 (pp. 2053–2058). https://doi.org/ 10.1109/GreenCom-iThings-CPSCom.2013.384. 20. Yassine, A., Singh, S., & Alamri, A. (2017). Mining human activity patterns from smart home big data for health care applications. IEEE Access, 5, 13131–13141. https://doi.org/10.1109/ ACCESS.2017.2719921 21. Gatouillat, A., Badr, Y., Massot, B., & Sejdic, E. (2018). Internet of medical things: A review of recent contributions dealing with cyber-physical systems in medicine. IEEE Internet of Things Journal, 5(5), 3810–3822. https://doi.org/10.1109/JIOT.2018.2849014

162

A. Sinha et al.

22. Dimitrov, D. V. (2016). Medical internet of things and big data in healthcare. Healthcare Informatics Research, 22(3), 156–163. https://doi.org/10.4258/hir.2016.22.3.156 23. Limaye, A., & Adegbija, T. (2018). HERMIT: A benchmark suite for the internet of medical things. IEEE Internet of Things Journal, 5(5), 4212–4222. https://doi.org/10.1109/ JIOT.2018.2849859 24. Harerimana, G., Jang, B., Kim, J. W., & Park, H. K. (2018). Health big data analytics: A technology survey. IEEE Access, 6, 65661–65678. https://doi.org/10.1109/ ACCESS.2018.2878254 25. Saheb, T., & Izadi, L. (2019). Paradigm of IoT big data analytics in the healthcare industry: A review of scientific literature and mapping of research trends. Telematics and Informatics, 41(March), 70–85. https://doi.org/10.1016/j.tele.2019.03.005 26. Mohanty, S., Mishra, A., & Saxena, A. (2021). Medical data analysis using machine learning with KNN. Advances in Intelligent Systems and Computing, 1166, 473–485. https://doi.org/ 10.1007/978-981-15-5148-2_42 27. Viceconti, M., Hunter, P., & Hose, R. (2015). Big data, big knowledge: Big data for personalized healthcare. IEEE Journal of Biomedical and Health Informatics, 19(4), 1209– 1215. https://doi.org/10.1109/JBHI.2015.2406883 28. Garg, N., Wazid, M., Das, A. K., Singh, D. P., Rodrigues, J. J. P. C., & Park, Y. (2020). BAKMP-IoMT: Design of blockchain enabled authenticated key management protocol for internet of medical things deployment. IEEE Access, 8, 95956–95977. https://doi.org/10.1109/ ACCESS.2020.2995917 29. Cao, R., Tang, Z., Liu, C., & Veeravalli, B. (2020). A scalable multicloud storage architecture for cloud-supported medical internet of things. IEEE Internet of Things Journal, 7(3), 1641– 1654. https://doi.org/10.1109/JIOT.2019.2946296 30. Parimi, S., & Chakraborty, S. (2020). Application of big data & iot on personalized healthcare services. International Journal of Scientific and Technology Research, 9(3), 1107–1111. 31. Kumari, A., Tanwar, S., Tyagi, S., Kumar, N., Maasberg, M., & Choo, K. K. R. (2018). Multimedia big data computing and Internet of Things applications: A taxonomy and process model. Journal of Network and Computer Applications, 124(October), 169–195. https:// doi.org/10.1016/j.jnca.2018.09.014 32. Kumar, M. (2020). A secure and efficient cloud-centric internet- of-medical-things-enabled smart healthcare. IEEE Internet of Things Journal, 7(10), 10650–10659. 33. Nazir, S., et al. (2020). A comprehensive analysis of healthcare big data management, analytics and scientific programming. IEEE Access, 8, 95714–95733. https://doi.org/10.1109/ ACCESS.2020.2995572 34. Kumar, S., & Singh, M. (2019). Big data analytics for healthcare industry: Impact, applications, and tools. Big Data Mining and Analytics, 2(1), 48–57. https://doi.org/10.26599/ BDMA.2018.9020031 35. Borovska, P. (2018). Big data analytics and internet of medical things make precision medicine a reality. International Journal of Internet of Things and Web Services, 3, 24–31, [Online]. Available: http://www.iaras.org/iaras/journals/ijitws 36. Supriya, M., & Deepa, A. (2020). Machine learning approach on healthcare big data: A review. Big Data and Information Analytics, 5(1), 58–75. https://doi.org/10.3934/bdia.2020005 37. Mangla, S. K., Raut, R., Narwane, V. S., Zhang, Z., & Priyadarshinee, P. (2020). Mediating effect of big data analytics on project performance of small and medium enterprises. Journal of Enterprise Information Management, 34(1), 168–198. https://doi.org/10.1108/JEIM-12-20190394 38. Yacchirema, D., De Puga, J. S., Palau, C., & Esteve, M. (2018). Fall detection system for elderly people using IoT and big data. Procedia Computer Science, 130, 603–610. https:// doi.org/10.1016/j.procs.2018.04.110 39. Hartley, J., Benington, J., Press, P., & Central, P. E. (2018). Leadership for healthcare. British Journal of Healthcare Management, 24(11), 548–550. 40. Sinha, A., Biswas, A., Raj, T., & Misra, A. (2020). Big data analytics on matrimonial data set. International Journal of Innovative Research in Applied Sciences and Engineering, 4(4), 722–728. https://doi.org/10.29027/ijirase.v4.i4.2020.722-728

Application of Big Data Analytics and Internet of Medical Things (IoMT). . .

163

41. Dave, D., Naik, H., Singhal, S., & Patel, P. (2020). Explainable AI meets healthcare: A study on heart disease dataset (pp. 1–23). https://doi.org/10.48550/arXiv.2011.03195 42. Ploug, T., & Holm, S. (2020). The four dimensions of contestable AI diagnostics - A patientcentric approach to explainable AI. Artificial Intelligence in Medicine, 107(January), 101901. https://doi.org/10.1016/j.artmed.2020.101901 43. Duan, Y., Edwards, J. S., & Dwivedi, Y. K. (2019). Artificial intelligence for decision making in the era of Big Data – evolution, challenges and research agenda. International Journal of Information Management, 48, 63–71. 44. Wang, X., Meng, X., & Yan, S. (2021). Deep learning-based image segmentation of cone-beam computed tomography images for oral lesion detection. Journal of Healthcare Engineering, 2021, 4603475. https://doi.org/10.1155/2021/4603475 45. Wu, Y., Shen, Y., & Sun, H. (2021). Intelligent algorithm-based analysis on ultrasound image characteristics of patients with lower extremity arteriosclerosis occlusion and its correlation with diabetic mellitus foot. Journal of Healthcare Engineering, 2021, 7758206. https://doi.org/ 10.1155/2021/7758206 46. Ahmad, M. A., Eckert, C., & Teredesai, A. (2018). Interpretable machine learning in healthcare. In Proceedings of the 2018 ACM international conference on bioinformatics, computational biology, and health informatics (pp. 559–560). 47. Pawar, U., O’Shea, D., Rea, S., & O’Reilly, R. (2020). Incorporating explainable artificial intelligence (XAI) to aid the understanding of machine learning in the healthcare domain. In AICS (pp. 169–180).

An Interpretable Environmental Sensing System with Unmanned Ground Vehicle for First Aid Detection Ali Topal, Mevlut Ersoy, Tuncay Yigit, and Utku Kose

1 Introduction Technological studies have progressed very rapidly in the last century. With the development of technology, terms such as “autonomous robots” and “Laser Imaging Detection and Ranging (LIDAR) technology” are frequently used today. In addition to performing simple, repetitive tasks quickly and precisely, today’s robots can quickly execute more complex algorithms, and the way they move and “think” can be compared to a human’s thinking structure. Robots, which make living standards easier every day, have become a guide by shedding light on new discoveries in human life. The advances in the technical field from the beginning of the century to the present have brought new inventions. As a result of these advances, new fields of work have emerged. Within the scope of the study, some of the more recent study topics were examined. Topics; autonomous robots can be classified as LIDAR and environmental sensing systems. Today, there are environment mapping systems developed for different purposes. These systems in which LIDAR technology is used are developed and studied in commercial or academic environments. However, current studies generally progress towards a single goal. We frequently encounter in our daily life; autonomous cars [1] that make daily driving a safer and more efficient experience by making decisions for the driver, autonomous cleaning robots that make the basic map of the environment with the LIDAR sensor [2] to be able to move without getting stuck with obstacles, and mapping of the environment in virtual reality applications with A. Topal () Department of Computer Technologies, Isparta Applied Sciences University, Isparta, Turkey e-mail: [email protected] M. Ersoy · T. Yigit · U. Kose Department of Computer Engineering, Suleyman Demirel University, Isparta, Turkey e-mail: [email protected]; [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_9

165

166

A. Topal et al.

the help of LIDAR sensors commercial or academic studies such as 3D modeling of archaeological monuments and places using LIDAR [3, 4], smart city planning using LIDAR data [5], plant detection and farmland modeling for agricultural robots using LIDAR sensor [6] have been carried out. Akay et al. [7] extracted the 3D characterization of forest areas with accurate Digital Elevation Modeling (DEM) using LIDAR remote sensing technology in order to avoid time loss and resource consumption in managing natural resources and collecting necessary information in large areas. In their study, Zhang and Singh [8] implemented real-time LIDAR audiometry and mapping system to avoid data loss in the point cloud created by laser scanners in motion. Wa¸sik et al. [9] developed a LIDAR-based relative positioning and tracking system. The presented approach is based on the use of two LIDARs positioned front and back to provide a full field of view to the autonomous robot they are using. Ka˘gızman [10] developed the 3D LIDAR system and made obstacle detection. For the coordinated operation of the LIDAR system, it used ROS, which is the software platform that enables the robot components to be controlled by a computer. In their study, Akyol and Ay¸segül [11] performed two-dimensional mapping of the environment and simultaneous position determination with the Simultaneous Localization and Mapping (SLAM) method using the LIDAR sensor. Perhaps the most significant disadvantage of using LIDAR sensors is their cost. Within the scope of the study, it is aimed to develop an environment modeling system with an unmanned ground vehicle (UGV) by using low-cost LIDAR and other low-cost components. Laser Imaging Detection and Ranging (LIDAR) sensors are devices that emit pulses of laser light to measure distance. It is similar to radar technology. Laser pulses are used instead of radio waves used in radar sensors. Data collected from sensors is used to create highly detailed 2D and 3D maps of the environments around us. While the traditional mapping method is used in fields such as surveying, agriculture, and mining, today, with the development of technology, LIDAR technology is preferred in many areas for more detailed mapping. With the help of complex algorithms, LIDAR sensors are used to solve the SLAM problem for use in autonomous navigation in military applications, autonomous vehicles, and urban search and rescue missions. Various SLAM algorithms are available for air, land, and sea platforms. SLAM is one of the preferred methods because it has many uses and can make much more precise measurements than GPS. There are two SLAM packages commonly used in ROS (Robot Operating System) software. These are GMapping and Hector SLAM. HectorSLAM method was used within the scope of the study. Perhaps the most critical disadvantage of using LIDAR sensors for SLAM is their cost. The study focused on the use of low-cost LIDAR. This study is aimed to create a 3D model of the place where it is located by facilitating access to the environments determined by the autonomous robot to be used in the areas where different mapping systems used in many areas are planned to be realized within the scope of the study, to realize the model at a more effective level and to be low cost. In detail, the system ensures an interpretable view in the context of IoT synergy, considering first aid detection at the time of disasters and any situation threatening human health.

An Interpretable Environmental Sensing System with Unmanned Ground. . .

167

2 Material and Method 2.1 Investigation of Unmanned Ground Vehicles Unmanned ground vehicle (UGV) is the general name for physically unmanned robots that can move on the ground. It is used in situations where it is dangerous or impossible for a human being to be present in the relevant environment. Today, unmanned ground vehicles are produced in many different sizes, shapes, configurations, and characters. Unmanned ground vehicles can be controlled remotely and perform their tasks autonomously within the framework of a certain action plan through the sensors used on them. The basic techniques of the unmanned ground vehicle can be classified as environment sensing, motion planning, chassis dynamics control, and cloud control. As a term, unmanned ground vehicle and autonomous ground vehicle are often confused. We can specify autonomous ground vehicles as a subset of unmanned ground vehicles. While unmanned ground vehicles can go on the determined route or the operational route by remote control, autonomous ground vehicles perform all their actions on the environment model they create by perceiving the environment and finding their location in this environment with the decisions made by artificial intelligence applications. The visuals of some UGVs developed today are shown in Fig. 1. Adding the autonomous control feature to a land vehicle can be expressed as modeling the relevant environment through sensors, finding the vehicle located within the scope of the model, and moving the vehicle by planning the vehicle movement with the applied decision and support algorithms. Architectures such as sensors, 2D-3D object finding and classification, distance measurement, sensor fusion, positioning, interpretation, decision, and control are used in the development of UGVs. Sensors are preferred for UGVs to describe their environment. Generally used sensors for land vehicles can be specified as radar, LIDAR, IMU, camera, and audiometry sensors. All sensors have properties that are good or bad for each other. In the environment and situation where some sensor works correctly, the other

Fig. 1 Example of unmanned ground vehicles [12, 13]

168

A. Topal et al.

Fig. 2 Preferability and performance of sensors

sensor may not work correctly. By testing such different sensors on vehicles and determining the most suitable sensor, with the help of sensor fusion algorithms, the disadvantages of different sensors between each other can be eliminated and optimum measurement can be taken. The graph about which sensors are preferable for different requirements used in autonomous systems is shown in Fig. 2.

2.2 Overview of Arduino Microcontroller and Raspberry Pi Embedded System for IoT Setup Arduino is a physical programming platform consisting of a development environment equipped with digital and analog input/output (I/O) pins and containing the processing/wiring language. The basic components of Arduino are Arduino integrated development environment (IDE), Arduino libraries, Arduino bootloader, AVRDude, and compiler (AVR-GCC). The development environment is written in Java and based on the processing language environment. Libraries used in programming are written in C and C++ languages and compiled with AVR-GCC and AVR Libc. As for the reasons for choosing Arduino, it has advantages such as being directly connected to the computer via universal serial bus (USB) and the availability of plug-ins and sensor circuit elements that can be easily adapted to the developed kits. It does not need the programmer required for microcontrollers. It offers more cost-effective and broad library support compared to similar systems. The use of the Arduino platform was preferred to build up the IoT perspective because all the operations performed for the control of the UGV developed within the scope of the study are carried out in a fast and stable manner (as it is vital for also first aid detection and immediate healthcare-oriented working mechanisms). When the characteristics of the Arduino boards were investigated, the Arduino Mega 2560 board was found suitable to be used within the scope of the project. Thanks to the expansion feature of the card, it can easily communicate with many peripheral units and different sensors, and work packages to be carried out within the scope of the

An Interpretable Environmental Sensing System with Unmanned Ground. . .

169

application can be carried out in a faster and more integrated manner. There are 54 inputs and outputs on the card. The card, which has 256KB RAM capacity, operates at a 16 MHz operating frequency and +5 V operating voltage. The other platform used in the study is the Raspberry Pi embedded system computer. It is a fully equipped embedded system computer, usually used with the Linux operating system, consisting of a single board with USB and Ethernet inputs on the RPI. By installing Linux or Windows operating systems, it can perform many operations that can be done with a normal computer. There is also the Raspbian operating system based on Debian, which is optimized for Linux-based RPI. This mini-computer, which was originally designed for educational purposes, is now used in many academic studies and commercial projects. Raspberry Pi 4 Model B was used in the system design because all operations performed for mapping and image processing processes in the developed UGV are carried out more efficiently. The card used has a quad-core 1.5 GHz 64-bit processor and 4GB of RAM. There are 40 input-output pins on the board.

2.3 Development of System Design Raspberry Pi and Arduino platforms were used together in order to obtain the most appropriate results and to ensure that the UGV works within the IoT environment at full performance. The data from the distance sensing sensors and the temperature– humidity sensor on the Arduino are transferred to the microcontroller. Motors are controlled by the motor driver card connected to the microcontroller. All these operations are transmitted to the server via the Wi-Fi module. Similarly, the data coming from the peripherals on the Raspberry Pi is transmitted to the server via the integrated Wi-Fi on the Raspberry Pi. In this process, the communication of two different platforms is provided by mobile applications and web interface applications. The interaction of the robot with the outside world, the observation, and control of the data obtained as a result of this interaction are monitored with a mobile application with a dynamic interface. The blog diagram of the developed IoT system is shown in Fig. 3. The steps carried out in the development process and the methods used are explained under separate headings.

2.4 Components Used in System Design In the system design developed, a triangular tracked chassis with high maneuverability, capable of moving in harsh environmental conditions, was used as the body design of the robot. There are two driving wheels and 10 bearing wheels on the chassis. The connection of the wheels is provided by plastic pallets measuring 4.5 cm × 78 cm. A protection frame has been designed to place the cards and sensors used in the development of the body design in the most suitable positions and to take

170

A. Topal et al.

Fig. 3 Blog diagram of the IoT setup

precautions against impacts and damages that may come from the surroundings. The general view of the used robot chassis and the parts it contains are shown in Fig. 4.

2.4.1

Arduino and Its Peripherals

Peripherals used for UGV’s motor control and acquiring ambient data are controlled by Arduino Mega. Pololu brand VNH5019 motor driver board is used on Arduino microcontroller to provide motor controls. Working between 5.5 V and 24 V, this board supports two bidirectional DC motor control. The maximum PWM frequency is 20 kHz. Another peripheral on the Arduino physical programming platform is the ultrasonic distance sensor. Four distance sensors are used in the autonomous driving

An Interpretable Environmental Sensing System with Unmanned Ground. . .

171

Fig. 4 Chassis overview and components

control of the UGV so that it can move forward without getting stuck with the surrounding obstacles. HC-SR04 Ultrasonic sensor, which can measure distances up to 4 m, was preferred within the scope of the study. It works with the principle of finding the distance by measuring the time it takes for sound waves to travel to and from objects. The developed land vehicle also collects temperature and humidity data of its environment. CJMCU-1080 HDC1080 sensor is used as a high precision temperature and humidity sensor integrated on Arduino. The relative humidity accuracy is ±2%, and the temperature accuracy is ±0.2 ◦ C [14]. All components used in system design are transferred wirelessly to the control point. ESP8266 WiFi module, which supports TCP/IP protocol, is used for wireless connection.

2.4.2

Raspberry Pi and Its Peripherals

LIDAR technology is used for the environment mapping system, which determines the main theme of the study. RPLIDAR A1M8 2D 360◦ laser scanner sensor was preferred for low-cost 3D modeling of the UGV’s environment. It is capable of 2000 samples per second at a scanning rate of 5.5 Hz. RPLIDAR’s default scanning rate can be adjusted by changing the PWM signal sent to the motor by the user between 2 and 10 Hz. Thus, a maximum sampling rate of 8000 per second can be achieved at a scanning rate of 10 Hz. The distance measurement is a maximum of 12 m. Its angular sensitivity in detection is ≤1◦ . The working principle of laser distance sensors is based on the calculation of the time taken until the laser beams they send hit the object and return in terms of distance [11]. Laser energy is transmitted to objects in a short time and at a certain distance. Any object within range of the laser sensor reflects a certain portion of the transmitted light energy. The returning light energy is detected by the detector of the laser sensor. The time it takes for the light energy from the sensor to hit the object and return is calculated. Since the speed of this light beam coming out of the laser, the speed of light is known, the distance between the object and the sensor can be easily found. How the laser scanner sensor works is shown in Fig. 5 simply.

172

A. Topal et al.

Fig. 5 Laser distance measurement

RPI Camera is used as another peripheral on the Raspberry Pi platform. Thanks to the two IR infrared LEDs on the camera, night vision is available and clear images are obtained in dark environments. The night vision recording distance of the camera is 8 m, and this camera with 5 MP resolution can record in 1080p sensor resolution.

2.5 Electronic Circuit Designs of the System The realization of robot control operations in the system design is provided by Arduino Mega 2560 and VNH5019 motor driver card connected to this card. The realization of the robot’s environmental sensing processes is provided via the Raspberry Pi 4 Model B system computer. A switch is placed between the Raspberry Pi and Arduino board fed by a 12 V lipo battery. In addition, a regulator with a voltage range of 0.8–20 V is positioned between the RPI-Arduino and the switch. Trig and echo pins of four HC-SR04 ultrasonic sensors with 5 V voltage are connected. The SDA and SCL pins of the high-precision temperature and humidity sensor operating with 3.3 V voltage are connected to the Arduino platform. The TX and RX pins of the ESP8266 sensor working with 3.3 V are connected. Finally, the connections of the LIDAR and the camera are made on the RPI. The electronic connection diagram of the components on Arduino and Raspberry Pi is given in Fig. 6. The connections of all peripherals used on the Arduino platform were provided by jumper cables during the development process. The printed circuit board is designed to reduce the physical dimensions and surface areas of the system design and to eliminate the cable complexity used in the circuit design. Arduino Mega 2560 board and all of the circuit elements connected to this platform have been transferred to the printed circuit board. As a result of the development of the printed circuit board, a more stable circuit design has been achieved.

An Interpretable Environmental Sensing System with Unmanned Ground. . .

173

Fig. 6 Arduino and Raspberry Pi electronic wiring diagram

2.6 Development of Arduino Control Software The real-time operating system FreeRTOS library was used in the coding process of the sensors used on the system. Thanks to this library, it is ensured that all processes are run simultaneously. For motor controls, DualVNH5019MotorShield library is used. The power connections of the motors are connected to the driver board with the M1A-M1B and M2A-M2B ports, and the PWM pins required for motor control are connected to the Arduino. It sends a maximum of 400 PWM signals to the Arduino motor driver. Speed and direction settings have been made for Motor 1 and Motor 2. Due to the characteristics of the driver card used, a value between −400 and 400 is given for the speed. The speed of the robot is set in three stages, fast (400), medium (200), and slow (50). When the motor is first started, it will act on the 200 PWM signal. Then, if desired, the speed of movement of the robot can be determined with the speed steps. The brake setting is set between 0 and 400. At this point, 400 corresponds to full braking. ESP8266 Wi-Fi module is used to provide remote control of Arduino and peripherals. I2C communication bus is used in the communication of this module. I2cMaster and Wire libraries are used in the operation of the ESP8266 module. Within the scope of the study, after power connections of four ultrasonic distance sensors with 5 V voltage used for autonomous control of motors, trig and echo pins were connected. Trig is the part of the sensor that sends the sound wave, and echo is the part that receives the sent sound wave. As the working principle of the sensor with the most efficient measuring range of 2–200 cm, it starts when a pulse of at least 10 µs (microsecond) duration is applied to the trig pin. In response, the sensor transmits a sonic burst of eight pulses at 40 KHz. These eight pulses come out with the device’s unique sound signature, allowing the echo to distinguish incoming sound waves from ambient noise. After eight ultrasonic sound pulses exit the trig and hit the object and reach the echo pin, the echo pin becomes HIGH to start generating the beginning of the signal. If the outgoing sound waves do not

174

A. Topal et al.

return, the echo signal will time out after 38 ms (millisecond) and decrease. Thus, an impact of 38 ms indicates that there is no obstacle in the sensor range. In this process, if the pulses hit the object and come back, the echo will be LOW as soon as the signal is received. This event generates a pulse that varies in width from 150 µs to 25 ms, depending on the time it takes for the signal to be received. The time elapsed between the departure and return time of the received signal is used to calculate the distance to the object. Temperature and humidity data are also used to recognize the environment in which the developed autonomous vehicle is located. At this stage, the MLX90615 library is loaded for the use of high precision CJMCU1080 HDC1080 temperature–humidity sensor with low power consumption.

2.7 Development of Embedded Interpretable Software View The development of the sensors used in the unmanned ground vehicle was carried out over the RPI. First, the Linux-based Raspbian operating system was installed on the system computer. SLAM algorithm on ROS was used for 2D mapping of the environment. The Melodic version of the Robot Operating System was installed according to the official ROS website instructions [15]. After this process, SLAM packages should now be installed. The rplidar_ros package, which receives the LIDAR data, and the SLAM package, which interprets this data to create maps, are installed from RoboPeak’s Github page [16]. Rviz visualization environment was used to display the 2D map as a result of scanning the environment where the robot is located. When the robot first starts up, Rviz opens with an incomplete map of the robot’s surroundings. The map is updated whenever the robot moves the distance specified by the map_update_distance_thresh parameter or rotates by an angle specified by the map_update_angle_thresh parameter. In the system design, besides the 2D map of the environment where the land vehicle is located, it also obtains an interpretable 3D model of the same environment with the developed software. The related software was developed in Python language. In the flow of the developed software, RPLidar and pygame modules were installed first, and the screen setup settings were made. The data has been scaled to fit the screen. In order to display the data coming from the laser scanner on the screen, a point cloud was created by determining the shape, size, and color properties. Obtained point clouds give only x and y axes due to the 2D laser scanner. The z-axis, which is required for 3D modeling, generates values in the most appropriate conditions determined in the software. The distance measurement accuracy for the z value of the LIDAR sensor, which has an angular sensitivity of ≤1◦ in detection, was determined as 0.5 mm. In addition, the LIDAR sensor on the robot performs scanning by moving up and down thanks to its adjustable platform. After the applied methods, the LIDAR sensor data running at a sampling rate of 8000 per second were combined in a file with the extension “.xyz” created in the software named lidar_scan.xyz. The file is saved in the main directory of the software. The sample coordinate output of the obtained point cloud data is given in Table 1.

An Interpretable Environmental Sensing System with Unmanned Ground. . . Table 1 Table captions should be placed above the tables

x −117.094867 1274.948479 1331.492504

175 y 383.000054 −594.518239 −592.818657

z 1.1 1.2 1.7

In addition to producing the 2D map and 3D model of the environment, an image processing technique is used to enable the vehicle to recognize the objects around it thanks to the camera. The software of the infrared RPI Camera used on the vehicle was developed with Python language. OpenCV open-source library is used in the software. With the camera used, the objects or living things around the vehicle were identified, and the data was transferred via the mobile application. At this stage, operations such as storing objects in the directory and assigning variables were carried out with the created camera.py file. The system was trained to identify and detect the objects with photographs of many different structures or shapes of the relevant object. Sample images of objects or living things are collected in the objects folder. The process of extracting the object or creature name from the image file name is also carried out in the camera.py file. The camera image is displayed in the interface designed with HTML language. The processes of running the code written in camera.py, transferring the camera to the interface, and displaying it in the developed mobile program are also developed in the main.py file. Raspberry Pi can be accessed with a remote desktop application to control the developed system and to make updates on the system in possible situations. VNC Server desktop sharing system, which works fast with Raspberry Pi, is used as a remote desktop application. In order to determine the optimum port in the remote connection settings, the configuration settings of the RPI have been made.

2.8 Development of Mobile Software and Interface Design Control of Arduino and Raspberry Pi platforms used in the development of UGV is provided via mobile application or desktop web application. In the desktop web application software, HTML and Java languages were used for interface design. NetBeans platform was preferred as the development environment. In addition, the mobile application has been developed over Android Studio. Map and camera buttons have been added to the application interface so that the image sizes are not reduced. The display of the relevant area is presented clearly by switching between the buttons. The direction buttons with motor control are positioned in the most appropriate place in the interface of the application in order to be accessible. The engine control speed of the land vehicle can be selected as one of three stages. In addition, the data of the temperature and humidity sensor on the Arduino platform can be displayed in the interface design. The general view of the interface application on mobile phones and desktop computers is shown in Fig. 7.

176

A. Topal et al.

Fig. 7 Interface application overview

With the developed interface application, an integrated interface design has been created where all the capabilities of the unmanned ground vehicle can be controlled, the environment map created by the LIDAR laser scanner can be viewed, the images created for a better definition of the environment with the camera using image processing technique can be observed, and other sensor data can be accessed on a single screen.

3 Research Findings 3.1 Interpretable Mapping Operations Outputs In the developed system design, a 2D LIDAR laser scanner was used for 2D mapping and 3D modeling of UGV’s environment. Operations within the scope of the study were carried out in an experimental environment specially prepared for land vehicles. Firstly, 2D mapping of the environment was performed with the LIDAR sensor positioned on the UGV. Since the LIDAR scanning height is limited in 2D mapping, the appearance of high-form objects on the map may be meaningless. In experimental studies, glazed objects such as windows encountered during mapping have caused significant complications. The laser light sent by LIDAR can penetrate glass objects instead of being reflected by LIDAR. These scenarios can often cause the SLAM algorithm to interpret these scans as objects too far away or not save the scans on the map at all, breaking the map with blurred or broken-looking walls. This can be seen in the area circled in red on the 2D environment map in Fig. 8. The three lengths, highlighted in yellow in Fig. 8, were compared with real-life measurements to determine how accurately the maps produced by the SLAM algorithm were scaled. Each square on the map represents 1 m. As a result of the comparison of the map measurements with the actual measurements, the first length was 5.9 m, the second length was 6.1 m, and

An Interpretable Environmental Sensing System with Unmanned Ground. . .

177

Fig. 8 Physical view and 2D map of the experimental environment

Fig. 9 Interpretable 3D model of the experimental environment

the third length was 0.9 m, and it was determined that they had the same values. The physical image of the experimental environment prepared within the scope of the study and the 2D environment map created is shown in Fig. 8. 3D models of the same environment are produced with the software developed on the Python platform. It is extracted in a 3D model of the same environment as the software developed on the Python platform. As with the mapping process, modeling glass objects can cause problems creating a point cloud due to missing data from Lidar. The more the LIDAR sensor scans the relevant environment with 8000 samples per second, the clear modeling of that environment is provided due to the density of the point cloud data. The point cloud created by the data obtained as a result of the scan is displayed in color based on the height value. The interpretable 3D environment model created in the experimental environment prepared within the scope of the study is shown in Fig. 9. Autonomous control of the land vehicle is monitored remotely with the interface application, enabling low-cost access to the camera and 2D-3D map data of the environment. The comparison of the developed system with some studies in the literature in terms of its features is given in Table 2.

178

A. Topal et al.

Table 2 Comparison of UGV with studies in the literature Study name Haala et al. [17] Weiss and Biber [6] Teixidó et al. [18] Wasik et al. [9] Ocando et al. [19] Ka˘gızman [10] Akyol and Ay¸segül [11] Developed UGV

Mapping 3D 3D 2D 2D 3D 3D 2D 3D

Auto. control X X X Yes Yes X Yes Yes

Interface There is X X Yes Yes X X Yes

Cost High High Low High High High Low Low

Fig. 10 Environment image with image processing technique

3.2 Additional Interpretability with Embedded System Camera To better recognize the environment in which the unmanned ground vehicle is located and to provide detailed interpretable information to the users, the image processing technique with the RPI camera was used. That is vital as the first aid detection by the system required that for finding people at the time of disasters such as an earthquake. The camera used on the embedded system computer also produces high-quality images in dark environments with its infrared feature. In identifying objects and living things, the system is trained with photos of the relevant object. As part of the study, about 10 sample images for each object or creature were sufficient to train the system. The identification process is stable for each object with the appropriate angle in the camera frame. Many objects can be defined in a single image frame. A cat trinket was also included in the environment so that it could be sampled in living detection at the experimental site. The camera image of the experimental environment prepared in the study and the camera image created by the image processing technique are shown in Fig. 10.

An Interpretable Environmental Sensing System with Unmanned Ground. . .

179

3.3 Developed UGV Design and Performance The control interface application of the developed system design and LIDAR mapping application are developed on the Raspberry Pi embedded system computer. On Raspberry Pi, the process of mapping the environment and transmitting the generated map to the mobile program works effectively and decisively. Because the embedded system computer had 4GB of RAM capacity, no slowdowns or problems were encountered at application runtime. In the design of the system, the engine controls of the land vehicle and the management of the peripheral units were provided with the Arduino Mega 2560 card. In the first phase of the project development process, the circuit design of all sensors and motors was provided with jumper cables while data flow problems occurred. The problem was solved by preparing a printed circuit board that collects all peripheral connections on a single card. It has been observed that radio waves coming out of Arduino and Raspberry platforms to transfer data to the server are transferred to the interface application without delay and data loss. The ultrasonic sensors on the UGV are located in the four corners of the chassis to ensure that the autonomous driving of the car takes place at high performance. The high-precision temperature and humidity sensor are located right in the middle of the chassis. LIDAR, which is used to model the environment in which the land vehicle is located, is positioned at the top of the chassis so that it can scan in optimal conditions, taking into account the size of the vehicle. As a result of this settlement plan, the system works decisively. The driving ability of the UGV has often resulted in poor mapping performance since it was at times below optimal. Most importantly, the vehicle often found it difficult to overcome smooth surfaces such as concrete floors. In order to minimize this problem and increase traction, a pallet is preferred made of ulpolen plastic material that does not break on impact around the wheels, is resistant to high temperature, and does not wear. It has been observed that the engines provide the necessary torque to move the land vehicle while the pallets grip the ground well. The developed UGV chassis provides the necessary performance and can perform its task decisively. The layout of the components on the developed UGV and the final design of the vehicle are shown in Fig. 11.

4 Discussion and Conclusions Currently, there are environment mapping systems developed for different purposes. These systems, which use LIDAR technology, are developed and worked in commercial or academic environments. But current studies are often developed for a single purpose. In experimental studies, there are no environmental sensor robot systems where LIDAR technology, camera with image processing technique, and other deconstructing sensors coexist. For this reason, it requires the use of different systems at high costs to get to know the relevant environment more. In addition,

180

A. Topal et al.

Fig. 11 Layout and final design of the components used in the UGV

throughout the developed system designs, the robot is controlled manually, or the data obtained by the robot cannot be transferred to the server in a stable manner. In research on mapping systems, sensors that usually make up the 3D data set were used as LIDAR sensors, and sensor control was often preferred by highly expensive freezing components such as the Nvidia Jetson TX Development Board. Furthermore, as mapping ensures interpretability in the context of IoT, it is still needed to develop informative interfaces for users directly. That is too important when first aid detection-based systems are developed for dealing with safe and interpretable unmanned technologies. In the developed system design, various peripheral units that will contribute to the definition of the environment are collected on a single robot together with LIDAR. Perhaps the most critical disadvantage of using LIDAR sensors is their cost. As part of the study, using low-cost LIDAR and other low-cost components, the autonomous robot facilitated access to designated environments and created a more effective 3D model of the space in which it is located. An integrated interface application has been developed where all the capabilities of the UGV can be controlled, the environment map created by Lidar can be displayed, the environment can be observed instantly with the camera for better identification of the environment, and other sensor data on the ground vehicle can be accessed on a single screen. For these studies developed on the Raspberry Pi card, a higher-performance embedded system computer that performs operations in a similar architecture can be used. Interface application, where UGV control is provided and data is displayed, communication with Arduino and Raspberry Pi over Wi-Fi cards via TCP/IP connection has been provided in a fast and stable manner. As part of the study, space scenarios that are not easy to use UGV were also considered. The vehicle must always run smoothly without sudden movements and aggressive turns due to collisions. In the chassis of the developed land vehicle, brass metal is preferred, and the durability of the vehicle skeleton is increased. The plastic material used in the pallets also provided a good grip on the ground of the car and the necessary torque performance of the engines.

An Interpretable Environmental Sensing System with Unmanned Ground. . .

181

As a result of the study, it was found that low-cost RPLIDAR is a suitable option for implementing the SLAM algorithm in environment recognition systems and environment modeling. This is very important as LIDAR was not an economical solution for many developers or students until recently. As this technology becomes increasingly available, progress and innovations can be made in areas such as autonomous navigation, cartography, search and rescue, and decommissioning industries. A project has been created to reduce the financial resources allocated to the materials used in work carried out by the institutions working on search and rescue or environment recognition systems with the decommissioned system. In addition, thanks to the developed UGV’s ability to control and monitor the processes in the environment, it is foreseen to increase human safety by using it in areas considered to be dangerous. 3D modeling of the relevant environment was provided with LIDAR, which created a 2D data set, and the data obtained was effectively displayed in an interpretable manner, allowing people and institutions to track the environment in detail. Thanks to these features, a portable, ergonomic, safe, integrated robot design has been created where the user can recognize the relevant environment by monitoring dangerous environments remotely. Thanks to the formed IoT synergy, the detection of people needing first aid in unreachable places can be ensured easily with an unmanned solution. That solution improves the use of the developed system for any type of disaster requiring people to reach out to difficult places and run unmanned solutions for interacting with other people in terms of healthcare purposes.

References 1. Pettigrew, S. (2018). [Online]. Available: https://theconversation.com/driverless-cars-reallydo-have-health-and-safety-benefits-if-only-people-knew-99370 2. Roborock. (2020). [Online]. Available: https://us.roborock.com/pages/roborock-s6-maxv 3. Stein, S. (2021). [Online]. Available: https://www.cnet.com/how-to/future-of-lidar-cool-nowgoing-to-be-cooler-apple-iphone-12-and-ipad-pro/ 4. Corns, A., & Shaw, R. (2009). High resolution 3-dimensional documentation of archaeological monuments & landscapes using airborne LIDAR. Journal of Cultural Heritage, 10, e72–e77. 5. Dwivedi, M., Uniyal, A., & Mohan, R. (2015). New horizons in planning smart cities using LIDAR technology. International Journal of Applied Remote Sensing and GIS (IJARSGIS), 1(2), 40–50. 6. Weiss, U., & Biber, P. (2011). Plant detection and mapping for agricultural robots using a 3D LIDAR sensor. Robotics and Autonomous Systems, 59(5), 265–273. 7. Akay, A. E., O˘guz, H., Karas, I. R., & Aruga, K. (2009). Using LIDAR technology in forestry activities. Environmental Monitoring and Assessment, 151(1), 117–125. 8. Zhang, J., & Singh, S. (2014, July). LOAM: Lidar Odometry and mapping in real-time. Robotics: Science and Systems, 2(9), 1–9. 9. Wasik, A., Ventura, R., Pereira, J. N., Lima, P. U., & Martinoli, A. (2016). Lidar-based relative position estimation and tracking for multi-robot systems. In Robot 2015: Second Iberian robotics conference (pp. 3–16). Springer. 10. Ka˘gızman, A. (2018). Otonom araçlar için 2B lazer tarayıcı kullanılarak yeni 3B LIDAR ˙ sistemi elde edilmesi ve engel tespiti. Istanbul Teknik Üniversitesi, Fen Bilimleri Enstitüsü, ˙ Yüksek Lisans Tezi, 100s, Istanbul.

182

A. Topal et al.

11. Akyol, S., & Ay¸segül, U. Ç. A. R. (2019). Rp-lidar ve mobil robot kullanılarak e¸s zamanlı konum belirleme ve haritalama. Fırat Üniversitesi Mühendislik Bilimleri Dergisi, 31(1), 137– 143. 12. Boston Dynamics. (2015). [Online]. Available: https://www.bostondynamics.com/legacy 13. Aselsan. (2017). [Online]. Available: https://www.aselsan.com.tr/ffdf27dd-4f67-4abb-bf029022a7ce042c.pdf 14. Texas Instruments. (2016). [Online]. Available: https://www.ti.com/lit/ds/symlink/hdc 1080.pdf?ts=1623524310036&ref_url=https%253A%252F%252Fwww.google.com%252F 15. ROS. (2018). [Online]. Available: http://wiki.ros.org/melodic/Installation/Debian 16. Github. (2018). [Online]. Available: https://github.com/robopeak/rplidar_ros 17. Haala, N., Peter, M., Kremer, J., & Hunter, G. (2008). Mobile LIDAR mapping for 3D point cloud collection in urban areas—A performance test. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 37, 1119–1127. 18. Teixidó, M., Pallejà, T., Font, D., Tresanchez, M., Moreno, J., & Palacín, J. (2012). Twodimensional radial laser scanning for circular marker detection and external mobile robot tracking. Sensors, 12(12), 16482–16497. 19. Ocando, M. G., Certad, N., Alvarado, S., & Terrones, Á. (2017, November). Autonomous 2D SLAM and 3D mapping of an environment using a single 2D LIDAR and ROS. In 2017 Latin American robotics symposium (LARS) and 2017 Brazilian symposium on robotics (SBR) (pp. 1–6). IEEE.

Impact of Pandemic Over Cognitive IoT for Healthcare Industry: A Market Study Deepshikha Bhargava

and Amitabh Bhargava

1 Introduction Internet of Things (IoT), the most promising technological innovation, stands for enabling communication among various interconnected objects or devices. The technical definition of IoT may vary according to the application in which it is required to implement. According to Oracle (What Is IoT n.d.), IoT is the interconnection of physical devices, sensors, software, and underlined technologies [1]. An IBM blog written by Clark (2020) defines IoT as the interconnection of devices and people for the purpose of communication and exchange of information, e.g. smart microwaves, self-driving cars, and smart wearables to name a few [2]. During the Red Hat Summit, IoT is all about connecting common household devices together, such as smart home objects, smart wearables, and medical devices [3]. According to Aeris, India (What Is IoT? Defining the Internet of Things (IoT) | Aeris 2020), IoT is defined as “The Internet of Things (IoT) refers to a system of interrelated, internet-connected objects that are able to collect and transfer data over a wireless network without human intervention” [4]. With the advent of digital transformation, IoT has become the prime technological solution reshaping the businesses, enterprises, and lives of individuals. The IoT interconnects all devices, objects, animals, and people beyond the intervention of humans/computers for the purpose of exchanging data over the internet. The

D. Bhargava () Amity School of Engineering and Technology, Amity University, Greater Noida, India A. Bhargava Amity Business School, Amity University, Greater Noida, India © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3_10

183

184

D. Bhargava and A. Bhargava

important fact to note is that there is no intervention of humans or computers in IoT [5]. A survey from Pawar et al. (2016) states various applications of IoT such as agriculture, personal care, healthcare, traffic monitoring, transport fleet management, hospitality, smart homes, smart grids in the energy sector, and water supply, to name a few [6, 7]. This chapter focuses on the applications of IoT in the healthcare domain.

2 IoT in Healthcare In the past decade, the emergence of technology advanced the healthcare sector. The technologies such as Big data, cloud, and IoT provide assisted remote patient monitoring, patient–doctor communication, and hospital management. IoT aided the interconnection of various patient monitoring and personal healthcare devices such as temperature monitors, glucometers, fatal monitors in ICUs, ventilators, electrocardiograms for heart patients to name a few. These IoT-enabled devices can record and transfer real-time patient health vitals to medical practitioners to monitor patients’ conditions more accurately and even from distant locations. Based upon the patient’s condition/reports from IoT devices, the doctors could plan follow-up interactions. However, early IoT solutions/applications were insufficient due to a lack of analysis. Further, artificial intelligence and data analytics aided IoT devices for better analysis of patients’ conditions. This amalgamation gave rise to smart IoT devices, smart wearables, and Internet Medical of Things (IoMT) and further significant healthcare progressions [8]. The IoT for healthcare is aimed at providing patient-centric, low-cost healthcare solutions, proactive and enhanced treatment, quicker disease diagnosis, better patient satisfaction and engagement, and remote monitoring. It contributes to safe healthcare solutions for patients, benefits their families and health insurance segment, and enables physicians and hospitals to deliver superlative care to patients, medical equipment, and drug management. However, the common concerns of healthcare IoT are the security and privacy of patient health data [9] (Fig. 1). The advancement in IoT application to healthcare includes the invention of various smart medical and patient-assistance devices. For example, smart wheelchairs in hospitals equipped with IoT sensors and monitoring software [1]; patient monitoring devices such as heartbeat monitors, nebulizers, ventilators and oxygen concentrators in ICUs or operation theatres; personal healthcare devices such as smart wearables, smart insulin pens, glucometers, health-hygiene and mood-monitoring devices [10], telemedicine and telehealth, connected imaging, public safety, and sportsmen care [11]. During the pandemic, healthcare demands less or no-contact solutions to avoid the risk of infections, such as remote patient monitoring, remote patient–doctor interaction, telehealth, secure transfer of electronic health record (EHR) and electronic medical record of the patients with hospitals/doctors. The advent of 5G

Impact of Pandemic Over Cognitive IoT for Healthcare Industry: A Market Study

185

Fig. 1 IoT for Healthcare contributions [8]

networks in IoT ensures secure, cost-effective, and safe transfer of medical data. During COVID-19, hospitals and clinics suffered from various challenges such as lack of oxygen, ventilators, beds, and other basic healthcare requirements. Through personal healthcare devices such as smart wearables, temperature monitoring, glucometers, heart rate monitors, and oximeters, the patient can record their vitals and further share them with their doctors. This not only reduces the unnecessary load on hospitals but also reduces the risk of spreading infections [8]. To establish better communication over smarter devices, the blend of cloud computing, artificial intelligence, and IoT helped in data analysis, data transfer, and data management over the secure cloud or mobile services. For example, the glucose level of diabetic patients is captured through a glucometer and uploaded to a cloud/mobile application. Further, the glucose level analysis was conducted to identify the adequate level of insulin requirement for the patients. Insulin pens further use this data so that the required level of insulin dose is injected into the patient using the smart device. Another example of this blend is Smart Nanny for elderly patients, where the smart camera records and analyzes all activities of the elderly, namely, their predefined routine, usual/well-defined route, and activity time. In case the elderly take more time in a specific activity or deviate from a route or sudden fall, recorded in real-time and further based upon analysis, the alert system activates. The IoT for healthcare now uses robotics/bots and virtual agents along with wearables for patient-centric care in terms of physical and mental health. These virtual agents/bots are enabled with voice commands, pill/doctor appointment and reminder management system, health and wellness tutorials, fall detection, recording patient vitals such as glucose and oxygen level, heartbeat, temperature, and electronic health records, to name a few. Hospitals and the healthcare industry also adopted IoT for hospital management, real-time inventory and staff management, and tracing of medical devices in hospitals [8].

186

D. Bhargava and A. Bhargava

Fig. 2 Healthcare IoT Architecture and Phases

IoT for healthcare first needs to integrate all IoT healthcare devices together to acquire data. Post data aggregation, the data from IoT devices in the analog form needs to convert into digital form. Further, the data is preprocessed, cleansed, and uploaded over the cloud. On this data, healthcare analytics provides an adequate result for better decision-making. This IoT for healthcare architecture follows four phases as mentioned in Fig. 2 [9].

3 IoT in the Healthcare Market As shown in Fig. 2, data integration and aggregation is the most vital phase of healthcare IoT; however, first-generation healthcare IoT lack integration and is yet to meet the next-generation healthcare IT market. Healthcare organizations are in search of suitable vendors dealing with healthcare delivery systems that can provide cost-effective, robustness, accessibility, and interoperability to patient data stored in EMR or HER [12]. Various healthcare segments include hospitals, clinics, surgical units, pharma, clinical R&D divisions, and Govt. agencies [11]. Jennifer Bresnick in the KLAS report states that “Healthcare organizations are actively seeking out enhanced data integration and interoperability features in a new generation of population health management tools.” According to this research on Healthcare IT trends, “PACS (Population Health Management)” is the priority of most healthcare organization due to its medical imaging and ERP tools services. Other promising vendors with high retention rates include “Innovaccer,” “HealthEC,” “Enli,” and “Epic Systems.” However, “IBM Watson Health Phytel” and “i2i Population Health” are on low priority due to their low customer service and retention rates. Other influential market player in healthcare IoT includes, “Philips,” “Lightbeam,” “athenahealth,” “Allscripts,” “Optum,” and new entrant “bevvy.”

Impact of Pandemic Over Cognitive IoT for Healthcare Industry: A Market Study

187

To qualify as a most promising and long-term vendor for a healthcare IoT ecosystem, the vendor needs to focus on high customer satisfaction, patient retention, strong integration capabilities, better patient experience, patient-centric approach, flexibility, streamlined IT-eco system, and proactiveness and responsiveness [12]. The IoT healthcare market segmentation includes IoT-based implanted or stationary medical devices, wearables, remote device management systems, software, and services. Market research on IoT in Healthcare Market (June 2020) identified the US, Ireland, Germany, Poland, Netherlands, Estonia, China, and the UK as key market players for IoT in healthcare. The report mentioned major key solution providers are in the US, which include Medtronic, Cisco Systems, IBM Corporation, GE Healthcare, Resideo Technologies, Agamatrix, Armis, Capsule Technologies, Intel, KORE Wireless, Microsoft Corporation, Oracle, OSP Labs, Oxagile, PTC, R-Style Labs, Sciencesoft, Softweb Solutions, STANLEY Healthcare, and Welch Allyn. Other key vendors are Comarch SA of Poland, SAP SE, Siemens and Bosch of Germany, HQSoftware of Estonia, Huawei of China, Royal Philips of Netherlands, and Telit of UK. This report also highlighted that the IoT in the healthcare market is anticipated to grow to 188.2 billion dollars by 2025, keeping in view the key drivers such as patient-centric healthcare, proactive and responsive patient engagement, and cost-effective healthcare solutions. Even the recent pandemic disruption has become a game player for IoT in the healthcare segment. Globally governments are investing in digital health and undertaking initiatives for promoting telehealth solutions keeping in view security and privacy. During the present pandemic, there is an opportunity to advance IoT healthcare applications such as smart wearables, health and wellness application, fitness bands, telehealth, remote patient monitoring, online consultation, and interactive pharmacy, to name a few [11].

4 Cognitive IoT in Healthcare Cognitive IoT is an infusion of AI-based logical reasoning to data acquired from IoT devices. It is considered a game changer during the present pandemic. In a traditional hospital setup, the doctors need to monitor and control the medication/treatment of patients. Cognitive IoT acts as a virtual healthcare assistant where the real-time patient vitals are compared with their electronic health record through cognitive IoT-enabled systems. Based on the analysis, the alerts are generated to notify doctors so that critical decisions can be taken to save the patient’s life. Cognitive IoT uses different technologies such as natural language processing (healthcare chatbots/virtual assistants); machine learning (analysis and predictions based upon healthcare data); image and video analytics (analytics based upon data provided by different sources like CCTV or patient’s medical imaging); and text analytics (analyze textual data from varied sources such as customer care, blogs, logs, tweets, etc.) [13].

188

D. Bhargava and A. Bhargava

The cognitive-IoT healthcare framework is comprised of smart medical devices, sensors, patients, healthcare professionals, and AI base intelligent system for predictive analytics. It includes the Cognitive Internet of Medical Things (Cognitive IoMT) [14]. For example, deep learning-based EEG analytics [15]; AI and ML enabled IBM’s smart clinical diagnosis platform; cognitive fabric to analyze physiological, behavioral, cognitive, and environment for physical and mental health [16]; and Convolution Neural network-based cognitive IoT for ECG data [17].

5 Key Drivers, Opportunities, Challenges, and Limitations There are several parameters of IoT in healthcare that impact healthcare delivery. The key parameters/drivers include age, lifestyle, personal health and hygiene, and patient vitals. The IoT healthcare has enablers as opportunities, such as govt. policy and strategies (especially during COVID), technological advancements, 5G network connectivity, 3GPP standards, and cyber security. At the same moment, there are significant challenges to address, such as acceptance, data analytics, safety, augmented intelligence, privacy and security, data storage and ownership, interoperability, and cost [18]. The need to provide smart products, smart equipment and devices and their maintenance, smart labels, and well-defined processes.

6 Impact of Pandemic over Cognitive IoT on Healthcare Market During the pandemic, the role of cognitive IoT has been significant in terms of providing remote patient monitoring, consultation, online medicine, telehealth and telemedicine, rapid RTPCR test at home/quarantine centers, data on medical platforms, contact tracing and real-time monitoring of COVID patients, screening and surveillance, and pandemic prevention and control [19].

7 Conclusion The market research in the IoT healthcare segment highlights the growth in this sector pre- and post-pandemic period. The various market players identified as key healthcare solution providers, mainly the US vendors. It is also advent that the blend of IoT with artificial intelligence and cloud, referred to as Cognitive-IoT, played a vital role in patient care, medical innovations, personal care, remote consultation, diagnosis, and clinical trials.

Impact of Pandemic Over Cognitive IoT for Healthcare Industry: A Market Study

189

Cognitive-IoT and Cognitive IoMT provides remotely accessible cost-effective solutions to all stakeholder such as governments, medical practitioners/doctors, patients, pharmacist, and hospitals. During the present pandemic, we have witnessed the significant role of Cognitive-IoT in providing noninvasive healthcare solutions, real-time patient data acquisition, quality of care, and remote physiological monitoring of patients. However, healthcare organizations are still struggling with next-generation solutions of Cognitive-IoT; the current pandemic provides a viable opportunity for this sector to grow.

References 1. What is IoT. (n.d.). Oracle India. Retrieved July 31, 2021, from https://www.oracle.com/in/ internet-of-things/what-is-iot/ 2. Clark, J. (2020, August 28). What is the internet of things, and how does it work? IBM Business Operations Blog. https://www.ibm.com/blogs/internet-of-things/what-is-the-iot/ 3. What is IoT? (n.d.). RedHat summit. Retrieved July 31, 2021, from https://www.redhat.com/ en/topics/internet-of-things/what-is-iot 4. What is IoT? Defining the Internet of Things (IoT) | Aeris. (2020, April 10). Aeris | India. https://www.aeris.com/in/what-is-iot/ 5. McClelland, C. (2021, February 4). What is the internet of things, or IoT? A simple explanation. IoT for all. https://www.iotforall.com/what-is-internet-of-things 6. Pawar, A. B., & Ghumbre, S. (2016, December). A survey on IoT applications, security challenges and counter measures. In 2016 international conference on computing, analytics and security trends (CAST) (pp. 294–299). IEEE. 7. The 9 most important applications of the Internet of Things (IoT). (n.d.). Fracttal USA Blog. Retrieved July 31, 2021, from https://www.fracttal.com/en/blog/the-9-most-importantapplications-of-the-internet-of-things 8. Chouffani, R. (2020, July 2). Future of IoT in healthcare brought into sharp focus. IoT Agenda. https://internetofthingsagenda.techtarget.com/feature/Can-we-expect-the-Internet-ofThings-in-healthcare 9. IoT in Healthcare Industry | IoT Applications in Healthcare – Wipro. (n.d.). Wipro. Retrieved August 1, 2021, from https://www.wipro.com/business-process/what-can-iot-do-forhealthcare-/ 10. Internet of Things (IoT) healthcare examples. (2020, December 2). Ordr. https://ordr.net/ article/iot-healthcare-examples/ 11. IoT in Healthcare Market. (2020, June). https://www.marketsandmarkets.com/ .https://www.marketsandmarkets.com/Market-Reports/iot-healthcare-market160082804.html?gclid=CjwKCAjw55-HBhAHEiwARMCszlIeugtCVBeEYWZo63PieGC4 Zr4vHnhRuvTW8VpZa9s0mMdgLn0eZRoCRDQQAvD_BwE 12. Bresnick, J. (2019, March 27). KLAS: Data integration needs Spur population health IT market. https://healthitanalytics.com. https://healthitanalytics.com/news/klas-data-integrationneeds-spur-population-health-it-market 13. Raman, A. (2016, June 8). Cognitive IoT – the game changer for India. ETHealthworld.Com. https://health.economictimes.indiatimes.com/news/health-it/cognitive-iot-the-game-changerfor-india/52659662 14. Hassanien, A. E., Khamparia, A., Gupta, D., Shankar, K., & Slowik, A. (2020). Cognitive internet of medical things for smart healthcare: Services and applications (Studies in Systems, Decision and Control) (Vol. 311, 1st ed., 2021st ed.). Springer.

190

D. Bhargava and A. Bhargava

15. Alhussein, M., Muhammad, G., Hossain, M. S., et al. (2018). Cognitive IoT-cloud integration for smart healthcare: Case study for epileptic seizure detection and monitoring. Mobile Networks and Applications, 23, 1624–1635. https://doi.org/10.1007/s11036-018-1113-0 16. IBM Watson. (n.d.). Cognitive IoT for Heathcare – IBM. https://researcher.watson.ibm.com/. Retrieved August 1, 2021, from https://researcher.watson.ibm.com/researcher/view_group. php?id=7866 17. Rani Roopha Devi, K. G., Mahendra Chozhan, R., & Murugesan, R. (2019). Cognitive IoT integration for smart healthcare: Case study for heart disease detection and monitoring. In 2019 international conference on recent advances in energy-efficient computing and communication (ICRAECC) (pp. 1–6). https://doi.org/10.1109/ICRAECC43874.2019.8995049 18. Kelly, J. T., Campbell, K. L., Gong, E., & Scuffham, P. (2020). The internet of things: Impact and implications for health care delivery. Journal of Medical Internet Research, 22(11), e20135. Published 2020 Nov 10. https://doi.org/10.2196/20135 19. Swayamsiddha, S., & Mohanty, C. (2020). Application of cognitive internet of medical things for COVID-19 pandemic. Diabetes and Metabolic Syndrome: Clinical Research and Reviews, 14(5), 911–915. https://doi.org/10.1016/j.dsx.2020.06.014

Index

A Actuators, 7, 119 ARIMA, viii, 25–54 Artificial intelligence (AI), vii–ix, 1–19, 103, 126, 137, 139–141, 144, 147, 148, 151, 154–156, 184, 185, 187, 188 Autocorrelations, 29 B Big data, viii, 59, 119, 129–160, 184 Big Data Analytics, viii, 52, 129–160, 184, 188 Biochemistry, 108 Biochemistry and hormone tests, 108 Body sensor network, 122–123 C Cholesterol data, viii, 103–116 Classification, viii, 9, 25, 61, 68, 69, 85, 91–100, 105–107, 133, 134, 139, 154, 167 Cloud computing, 1, 57, 185 Cognitive data, 187 Cognitive IoT, vii, viii, ix, 119–128, 183–189 Comparative evaluation, 158–160 Computational intelligence, 31, 44 COVID-19 (CV-19), viii, ix 8, 10, 25–54, 73–88, 119, 185 COVID-19 pandemic, viii, 43, 73, 75–80, 183–189 Cyber security, viii, 188

D Data analysis, 52, 54, 57–70, 103, 121, 141, 153, 185 Data analytics, viii, 51, 129–160, 184, 187, 188 Database, 5, 10, 92, 104, 143 Data cloud, 7, 184 Data computation, 108, 139 Data interpretability, vii, 156, 159, 180 Data mining, 131, 133, 139, 140, 152, 166 Decision tree, 26, 97–99, 104–113 Deep learning (DL), vii, viii, 2, 3, 5–8, 14, 16, 44, 65, 67–70, 73–88, 154, 188 Deep neural network (DNN), viii, 5, 6, 84, 91–100 Detection, viii, ix, 1, 14, 57, 83, 85, 91–100, 119, 165–181, 185 Diagnosis, viii, 1, 5–7, 75, 91, 135, 141, 154–157, 160, 184, 188

E Economical analysis, 129, 136 Environmental sensing, viii, 165–181 Explainability, 2, 11–19, 155 Explainable AI (XAI), viii, 1–19, 91, 98, 99, 103, 104, 114, 154–160

F First aid, viii, 165–181 Future pandemic, 54

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 U. Kose et al. (eds.), Interpretable Cognitive Internet of Things for Healthcare, Internet of Things, https://doi.org/10.1007/978-3-031-08637-3

191

192 G Generalized linear model (GLM), 104–105, 108, 109, 112, 116 Geospatial, viii, 25–54 Gradient boosted trees, 104, 107, 110–112, 114–116

H Healthcare, 1, 26, 57, 77, 91, 119, 130, 168, 186 Healthcare information, 142 Hormone, 108 Hybrid machine learning, 60

I Image analysis, 4 Image data, 92, 93, 96 Image processing, 169, 176, 178, 179 Image segmentation, 84, 85, 92, 93 India, viii, 25–54, 147, 183 Industry analysis, 185 Intelligent systems, vii, 188 Internet of Healthcare Things (IoHT), viii, ix, 1–19, 91, 93, 98, 99, 104, 114 Internet of medical things (IoMT), viii, 119, 129–160, 184, 188, 189 Internet of Things (IoT), 1, 52, 57, 104, 119, 130, 166, 185 Interpretability, vii, viii, 2, 26, 155, 156, 159, 160, 178, 180 Interpretable artificial intelligence, viii, 2, 155 Interpretable machine learning, viii, 52, 103–116, 157–160 Interpretable sensing, 165–181

K Knee x-ray, viii, 91–100

L Laser Imaging Detection and Ranging (LIDAR), 165–167, 171, 172, 174, 176, 177, 179–181

M Machine learning (ML), viii, 3, 5, 9, 13, 15, 25, 52, 62, 67, 68, 84, 103–116, 126, 131, 133, 144, 147–148, 154–160, 187, 188 Market study, ix, 183–189 Massive data, 5, 7, 51, 138, 139, 143

Index Medical, 2, 58, 76, 91, 104, 119, 129, 184 Medical applications, 105, 129–160 Medical data, 4, 5, 15, 132, 140, 141, 185 Medical detection, 1, 14, 83, 91–100, 119, 165–181 Medical diagnosis, viii, 1, 5–7, 75, 91, 135, 141, 154–157, 160, 184, 188 Medical images, viii, 4, 6, 7, 9, 10, 14, 83–85, 91–100, 104, 138, 141, 144, 148, 156, 169, 172, 175–179, 187 Medical internet of things, 119, 131, 159 Medical precautions, 53, 78 Medical treatment, 15 Monitoring, 1, 8, 57, 63, 120–122, 125, 126, 130, 133, 134, 181, 184, 185, 187–189 Multi-party computation, viii, 57–70

N Network systems, 57, 119, 120, 125

O Osteoarthritis (OA), viii, 91–100

P Pandemics, viii, ix, 43, 54, 73, 75–80, 88, 183–189 Particle swarm optimization, viii, 91–100 Patient tracking, ix Personal data, 15, 58, 59, 70, 147, 154 Personal healthcare, viii, 119–128 Prediction, viii, 1, 2, 4–6, 8, 11–14, 17, 25–54, 59, 60, 65–70, 85, 103–116, 133, 155–158, 160, 187 Predictive analysis, 160, 188 Privacy preserving, viii, 2, 3, 57–70 Public healthcare, 14, 63, 140

R Real-time videos, viii, 73–87 Responsibility, 15, 18, 19, 45, 130, 136 Responsive AI, 2, 3 Robotic systems, viii, 165–167, 169, 170, 172–174, 179–181, 185

S Secure computation, viii, 57–70 Security, viii, 7–9, 58–60, 79, 120, 121, 124, 126, 127, 138, 139, 142–144, 158, 160, 184, 187, 188

Index Segmentation, 83–85, 92–94, 100, 187 Sensitive data, viii, 7, 8, 57, 59, 70, 127, 135 Sensor networks, 57, 119, 120 Sensors, 7, 57, 58, 63, 104, 119–126, 129, 131, 132, 136, 141, 148, 150–152, 165–177, 179, 180, 183, 184, 188 Simultaneous Localization and Mapping (SLAM), 166, 174, 176, 181 Smart applications, 125 Smart healthcare, viii, 1–19, 121, 124, 127 Smart tools, vii Social distancing, viii, 73–88 Statistical data, 52 Survey, viii, 4, 59, 119–160, 166, 184 System, 1, 26, 57, 77, 98, 103, 119, 131, 165, 183

193 T Telemedicine, 134, 184, 188 Tracking, ix, 59, 75, 119, 166

U Unmanned ground vehicles (UGVs), viii, 165–181

V VLDL cholesterol, viii, 103–116

X X-ray images, viii, 10, 14, 91–100, 141