Our bodies are definitely the most precious possession of our lives. Most people are concerned about their bodies; and thus spend a great deal of time and resources to maintain it at its best possible condition. As a consequence, the volume of data generated from health monitoring devices is increasing very rapidly, particularly through the use of wearable devices that we use to track our health, fitness, and well-being on daily basis [1]. Increasingly, the use of these informal devices in the outpatient context is being incorporated to provide a more holistic assessment of individuals. In general, these datasets are complex and encapsulate physiological and biomechanical information collected using arrangements of different sensors contained in a single wearable device such as a digital wristwatch [2].

An important reason why such massive and diverse datasets are required to describe, monitor, and diagnose the status or condition of persons is the inherent complexities of our bodies. While this data is the source for comprehensive assessment of people's state of health, the job of analysing and interpreting this data is performed by specialists e.g. clinicians and personal trainers. This is often a challenging job and makes a compelling case for automatic analysis and interpretation. Even partial but effective automation can be of great help to physicians, not only allowing laborious and error prone tasks to be carried out more efficiently and accurately, but also providing specialists with the ability to focus on other tasks that truly require human intervention, judgement, and expertise.

While clinical decision support systems are clearly beneficial, their design and development pose a challenging problem. The latter is associated with the volume and types of data involved, as well as the fact that clinical decisions are often predicated on a multiplicity of factors e.g. physiological conditions, comorbidities, and individuals' demographics. Furthermore, in some cases, the underlying pathophysiology behind the illnesses and conditions are not fully understood, and varies from patient to patient. As such, traditional top-down and rule-based systems are not always effective frameworks for implementing such a tool. What is required are flexible data driven frameworks that can be applied to different types of data and problems. In simple terms, there is a constant need for intelligent systems capable of generalising and adjusting to changes dictated by new incoming data. The degrees of freedom within the framework should also be adjustable to cope with the varying complexity of the problem.

Why are we writing this book? Simply because we believe Machine Learning (ML) is an important ingredient in the solution as it overcomes many of the limitations of traditional frameworks. Our intention is to demonstrate its potential by showcasing some recent advances. We feel incredibly fortunate that several world class researchers and engineers have agreed to share their work in this book. We hope that it will inspire our reader whatever their background may be.

Following on the previous chapter of machine learning on the previous chapter of machine learning (ML) advancements for optimisation and planning in global health policy decisions, here we discuss the mechanisms that enable such an approach through the global sharing of data, simulations and compute. This has been done through both institutional and crowd-sourced participation in global health challenges. Motivating a platform which enables trusted contributions and coordination of shared resources. Specifically demonstrating results applied to malaria endemic countries and the COVID-19 pandemic, and how such approaches may be extended towards future epidemic preparedness. Through this work, we aim to envision new approaches and systems to answer future global health challenges.

Computational measurements and interpretation of consciousness and unconsciousness states have evolved over the years. The computational representational understanding of mind hypothesis has yet to be tested as there are still multiple surrogate variables used for understanding how the CNS structures process data to generate outcomes. ML helps to measure and test physiological signals. However, the outcome is subject to interpretation in context. Nevertheless, discrimination between physiological states, whether these are healthy states, neurological disorders or pharmacologically induced states, seems to be better understood with the use of ML algorithms, but a multimodal approach to acquire neurological outcome is required for more objective measurement of neurophysiological responses. The field of ML of neurocognitive science is transformative and has great potential, but is still in its embryonic phase.

Cardiac arrhythmias are disorders in the electrical activity of the heart, mainly characterized by fast or slow heartbeats and often accompanied by an irregular rhythmic pattern. Some arrhythmias also exhibit morphological changes in the electrocardiogram (ECG) because of damage of heart cells, resulting in abnormal conduction of the impulses through different pathways. While some arrhythmias are benign, others are life-threatening or may lead to severe health complications.

Automatic arrhythmia detection is not something novel; and in fact, most of medical bedside monitors and cardiac telemetry systems these days have embedded capabilities for detecting a wide range of these. Algorithms for arrhythmia detection have been developed using different kinds and even combinations of input data - e.g. photoplethysmography (PPG), seismocardiograms (ECG), ballistocardiograms (BCG), and electrocardiograms (ECG). The latter is the oldest, most popular, and perhaps the most reliable data source - as it directly represents the electro-chemical conductive properties of the heart.

Unfortunately, ECG signals can be easily corrupted by motion artefacts resulting from locomotion or other forms of physical activity. Furthermore, physical activity results in increased heart rate, introducing more confounding factors into the problem - for example, some forms of tachycardia are characterized by high heart rates as those seen during exercise. Even worse, high heart rates shorten the distance between heartbeats - which makes the detection of irregular rhythms more difficult. Moreover, some arrhythmias share similar characteristics among them or with noisy signals, adding more complications into the equation. Thus, despite the advances in commercial wireless wearable/portable systems and apps for arrhythmia detection (e.g., KardiaMobile https://www.alivecor.com), the task of arrhythmia detection can still be challenging.

While all arrhythmias are worth of research, atrial fibrillation (AF) is of particular interest. The rationale for this is twofold: first, this is the most common arrhythmia with the highest incidence and prevalence worldwide. According to the British Heart Foundation [1] 110,102 patients were admitted in hospitals in the UK due to AF as the main diagnosis. In addition, data from the Framingham Heart Study [2] showed a three-fold increase in the prevalence of AF over the last 50 years; whereas the Global Burden of Disease report [3] estimated a worldwide prevalence of more than 40 millions of patients in 2016 [4]. Second, AF is a life-threatening arrhythmia and a lead cause of stroke, which in turn often results in dead or permanent disabilities. Therefore, the central focus of this chapter is on presenting a comprehensive review of advances in algorithmic development for detection of AF from ECG signals in the last three decades; as well as providing the reader with our own experience on the development of different methods to detect this condition based on aspects such as data availability and labeling.

Artificial Intelligence (AI) and machine learning (ML) have unarguably become fashionable terms within the discourse of healthcare research in the past decade. Like many novel technologies that have come to past, the hype is not substantiated by immediate success. However, AI/ML has made steady progress in clinical research. The area of human performance is not immune to the appeal of this technology. Like other industries, the question remains unanswered as to what extent AI/ML can be practically employed to enhance physical performance outcomes.

In recent years, there is a growing interest in using ML to model physical adaptations to training and predicting athletic performances, all of which will be further discussed. In these cases, ML was applied to transform raw physiological data to advanced, interpreted information based on which athletic coaches make informed decisions. However, the challenge remains to harness the true potential of this technology to impact physical performance of individuals as well as athletes in a measurable, verifiable, and scalable manner.

There is a need to bridge the gap between data scientists and human performance experts in order to translate the enthusiasm in AI into material transformation of physical performance. This chapter endeavours to do so by presenting the tenets of physical performance in ways that are intuitive for data scientists, such that they can find footholds to employ ML applications on problems that can enable improvement in physical performance. Vice versa, by suggesting steps to apply ML in physical performance, the chapter hopes to lower the barrier for sports and exercise experts to integrate data-centric approaches to training. Ultimately, the aim is not only to aid in realising the goal of "higher, faster, and stronger" in competitive sports but also to improve quality of life through physical activity.

We ventured out to explore the role of machine learning (ML) in the medical technology industry and were rewarded with a fascinating tapestry that reveals the rich and layered nature of these applications. Indeed, even in this relatively short book, we saw ML applied to a large-scale problem - predicting the mortality rate of malaria as a result of policy changes in Chapter 2, to microscopic ISFET based neural networks used to analyse sweat contents directly on the skin surface in Chapter 10, while not forgetting the many exciting applications in between.

It is undisputed that ML is a powerful technique that has been applied in many areas of the industry. In tandem, exciting research is on-going in many areas, which can provide novel and valuable tools for clinicians and improve the level of patient care. In spite of these benefits, the uptake of ML in the industry has been slow partly because of the heavily regulated nature of the medical devices industry as well as the fact that there is considerable hype around ML. Unfortunately, this is accompanied with unrealistic expectations and the subsequent disappointments have affected the standing of ML as a serious tool that can be used to improve health outcomes. Apart from inspiring readers about the possibilities of ML, we hope that this book can help with clarifying what ML means and correct any misconceptions about it.

In this chapter, we will start by reviewing the preceding chapters of this book, beginning with the types of data that are used, followed by the ML frameworks and methods that are used. Clearly, effective implementation and deployment of ML is crucial for its success and there are a wide range of factors involved. We will focus solely on the types of platforms where ML is deployed and highlight some interesting methods that are used to overcome challenging conditions where ML is deployed. Since it is important for manufacturers to demonstrate regulatory compliance, we will discuss some of the issues that they face during the ML design and deployment phase. Finally, we will conclude the book with some thoughts about the future of ML in healthcare.