Applications of Machine Learning in Digital Healthcare (Healthcare Technologies)

HEALTHCARE TECHNOLOGIES Cover 1 Contents 6 About the editors 14 1 Introduction 16 1.1 Why? 16 1.2 How? 17 1.3 What is ML? 17 1.4 The problem 18 1.5 Gradient descent 20 1.6 Structural components of the ANN 22 1.6.1 The fully connected neural network 22 1.6.2 Convolutional neural network 30 1.6.3 Pooling layers 34 1.6.4 The SoftMax function 36 1.6.5 Putting them together 39 1.7 Training and evaluating a neural network 40 1.7.1 Data organisation 40 1.7.2 Types of errors and useful evaluation metrics 41 1.7.3 ADAM optimisation for bias reduction 47 1.7.4 Regularisation for variance reduction 49 1.8 Conclusion 50 References 50 2 Health system planning and optimisation – advancements in the application of machine learning to policy decisions in global health 52 2.1 Model-based decision making 52 2.2 ML surrogates for prediction from epidemiological models 55 2.2.1 Gaussian process regression 56 2.2.2 Action-value function example 58 2.2.3 Epidemiological model calibration 61 2.2.4 Bayesian optimisation 62 2.3 Online learning 63 2.3.1 Stochastic multi-armed Bandit 64 2.4 Running epidemiological simulations as Bandits 66 2.4.1 Time 66 2.4.2 State 67 2.4.3 Action 67 2.4.4 Reward 68 2.4.5 Bandit approaches for simulated learning 70 2.4.6 Extensions to online learning 71 2.5 Reinforcement learning 72 2.5.1 State 74 2.5.2 Action 75 2.5.3 Reward 75 2.5.4 Markov decision processes 76 2.5.5 Cumulated return 76 2.5.6 Policy 77 2.5.7 Value function 77 2.5.8 Partially observable MDP (POMDP) 79 2.5.9 Learning sequential surrogate models from episodic simulators 79 2.5.10 Prediction – learning a value function 80 2.5.11 Simulation-based search – decision trees 81 2.5.12 Monte Carlo tree search (MCTS) 83 2.5.13 Gaussian process regression with selection inMCTS for learning sequential surrogates (GP–MCTS) 84 2.6 Control – optimal policies and planning 87 2.6.1 Optimal policy learning 87 2.7 Comparing predictions from multi-step and one-step methods with direct experience 88 References 89 3 Health system preparedness – coordination and sharing of computation, models and data 94 3.1 Computation 94 3.1.1 A proposed infrastructure 95 3.1.2 Platform components 97 3.1.3 Performance results 98 3.1.4 Example: technical approach for competitions 98 3.1.5 Environment web service 99 3.1.6 Competition API 99 3.1.7 Example code 100 3.1.8 Related work 102 3.2 ML competitions for health system preparedness 102 3.3 Planning from learnt models 104 3.4 KDD Cup 2019 and other competitions 105 3.4.1 Evaluation framework 106 3.4.2 Submission and scoring 109 3.4.3 Other competitions 109 3.5 Collaboration from competition 113 3.6 Example: analysis of successful competition approaches 113 3.6.1 Conclusions on competitions for health system planning 115 3.6.2 Human-in-the-loop 116 References 116 4 Applications of machine learning for image-guided microsurgery 122 4.1 Preoperative data collection 123 4.2 Preprocessing 125 4.2.1 Intensity histograms 126 4.2.2 Noise reduction 127 4.2.3 Contrast adjustment 135 4.2.4 Preprocessing review 143 4.3 Segmentation 143 4.3.1 Thresholding 144 4.3.2 Region-based thresholding 144 4.3.3 Edge-based thresholding 152 4.3.4 Post-processing 155 4.3.5 Validation 156 4.4 Registration 157 4.4.1 Image labeling 157 4.4.2 Feature identification 158 4.4.3 Feature matching 163 4.4.4 Transformation 163 4.5 Visualization 167 4.5.1 Real-time motion tracking 167 4.5.2 Overlaying 167 4.5.3 Image-guided microscopic surgery system 168 4.5.4 Augmented-reality-based microsurgical systems 168 4.6 Challenges 169 4.6.1 Infrastructure challenges 170 4.6.2 Safety challenges 170 4.6.3 Cost challenges 170 4.7 Chapter review 171 References 171 5 Electrophysiology and consciousness: a review 178 5.1 Introduction 178 5.2 Nervous system signals 179 5.2.1 Central nervous system 179 5.2.2 ANS 180 5.2.3 CNS–ANS connection in physiological mechanisms 182 5.3 Neurophysiological signal recording 183 5.3.1 Recording the electroencephalogram (EEG) 184 5.3.2 Recording the ECG 186 5.4 Applications of biopotentials in health and disease 188 5.4.1 Neurodegeneration 189 5.4.2 Anesthesia 189 5.4.3 Peri-operative stress 192 5.5 Analysis tools 193 5.5.1 ECG analysis 193 5.5.2 EEG analysis methods 196 5.5.3 Machine learning methods 201 5.6 Conclusion 202 References 203 6 Brain networking and early diagnosis of Alzheimer's disease with machine learning 212 6.1 Background 212 6.1.1 A brief history of brain study 212 6.1.2 Modern understanding of the brain 213 6.2 Laboratory model of brain connectivity 214 6.3 Problem definition 215 6.4 Devices used in AD diagnosis 216 6.5 Data types 217 6.6 Data preprocessing of MRI data 219 6.6.1 Median filters 219 6.6.2 Physiological noise removal by means of deconvolution 222 6.6.3 Image fusion 224 6.7 Machine learning for early AD diagnosis 230 6.7.1 SVMs 231 6.7.2 Deep learning 232 6.7.3 SVM techniques 233 6.7.4 Deep learning techniques 236 6.8 Conclusion 239 References 240 7 From classic machine learning to deep learning advances in atrial fibrillation detection 244 7.1 Physiology essentials 245 7.1.1 The healthy heart 245 7.1.2 Atrial fibrillation 245 7.2 Detection of AF 246 7.2.1 AF detection based on beat-to-beat irregularities 247 7.2.2 AF detection based on the ECG waveform morphology and hybrid methods 257 7.3 Conclusions 267 References 269 8 Dictionary learning techniques for left ventricle (LV) analysis and fibrosis detection in cardiac magnetic resonance imaging (MRI) 274 8.1 Introduction 274 8.2 Basics of dictionary learning 275 8.2.1 Probabilistic methods 276 8.2.2 Clustering-based methods 278 8.2.3 Parametric training methods 278 8.3 DL in medical imaging – fibrosis detection in cardiac MRI 278 8.4 HCM and fibrosis 279 8.4.1 Myocardial fibrosis in HCM 280 8.5 Cardiac magnetic resonance imaging with LGE-MRI 281 8.6 The assessment of cardiac fibrosis detection in LGE-MRI: a brief state-of-the-art 282 8.7 The proposed method 284 8.7.1 Feature extraction 285 8.7.2 Clustering 287 8.7.3 DL-based classification: training stage 287 8.7.4 DL-based classification: testing stage 288 8.8 First experiments and results 288 8.8.1 Study population 288 8.8.2 Results 289 8.8.3 Evaluation 291 8.9 Qualification and quantification of myocardial fibrosis: a first proposal 292 8.10 Conclusion 296 References 296 9 Enhancing physical performance with machine learning 302 9.1 Introduction 302 9.2 Physical performance and data science 303 9.2.1 Physical performance overview 303 9.2.2 The role of data in physical performance 303 9.2.3 Why ML? 305 9.3 Contextualise physical performance factors: ML perspectives 309 9.3.1 Training 309 9.3.2 Nutrition 312 9.3.3 Sleep and recovery 315 9.4 ML modelling for physical performance problems 317 9.4.1 Choosing ML models for the right physical performance tasks 317 9.4.2 Contributing ML features and methods 319 9.4.3 Challenges 321 9.5 Limitation 322 9.6 Conclusion 323 References 323 10 Wearable electrochemical sensors and machine learning for real-time sweat analysis 332 10.1 Electrochemical sensors: the next generation of wearables 332 10.2 The mechanisms and content of sweat 334 10.3 Considerations for on-body sweat analysis 336 10.3.1 Sweat gland densities and sweat rates 336 10.3.2 Sweat collection techniques and challenges 337 10.4 Current trends in wearable electrochemical sweat sensors 338 10.4.1 Common features of wearable sweat sensors 339 10.4.2 Opportunities for ISFETs and machine learning in wearable sweat sensing 341 10.5 The ion-sensitive field-effect transistor 342 10.5.1 The fundamental theory of ISFETs 343 10.5.2 ISFETs in CMOS 345 10.5.3 ISFETs in CMOS for sweat sensing 346 10.5.4 Existing ISFET-based wearable sweat sensors 349 10.6 Applications of machine learning in wearable electrochemical sensors 351 10.6.1 Existing research into ML for biosensors 351 10.6.2 Existing research into ML for ISFETs 351 10.6.3 Integration of analogue classifiers with ISFETs in CMOS 352 10.7 Summary and conclusions 359 Acknowledgements 359 References 359 11 Last words 368 11.1 Introduction 368 11.2 A review of the state-of-the-art 369 11.3 Implementation and deployment 371 11.3.1 Traditional computing and the memory hierarchy 372 11.3.2 Graphics processing unit 375 11.3.3 Hardware accelerators 380 11.4 Regulatory landscape 384 11.4.1 A brief interlude 385 11.4.2 Software development life cycle 387 11.4.3 Risk management in medical software development 393 11.4.4 Challenges specific to ML 400 11.5 Conclusion 403 References 403 Index 406 Back Cover 416 HEALTHCARE,TECHNOLOGIES Machine learning algorithms are increasingly finding applications in the healthcare sector. Whether assisting a clinician to process an individual patient's data or helping administrators view hospital bed turnover, the volume and complexity of healthcare data is a compelling reason for the development of machine learning based tools to aid in its interpretation and use. This edited book focuses on the applications of machine learning in the healthcare sector, both at the macro-level for guiding policy decisions, and at the granular level, showing how ML techniques can be applied to process an individual patient's medical data to swiftly aid diagnosis. Written by an international team of experts, the book presents several applications of machine learning in the healthcare sector, including health system planning, optimisation and preparedness, outlining the benefits and challenges of coordination and data sharing. Machine learning has many applications in processing patient data and topics such as arrhythmia detection, image-guided microsurgery and early detection of Alzheimer's disease are discussed in depth. The book also looks at machine learning applications exploiting wearable sensors for real-time analysis and concepts around enhancing physical performance. Suitable for an audience of computer scientists, healthcare engineers and those involved with digital medicine, this book brings together a plethora of machine learning applications from across the board of the healthcare services.