Model-Based Electrodermal Activity Sleep Events Detection Algorithm

Abstract

Despite the pivotal role of sleep in one’s life, quantitative methods to investigate it have been lagging behind. Recent years have witnessed a spike in the application of machine learning (ML) to automate the analysis of physiological signals. Throughout this dissertation, we focus on the electrodermal activity (EDA) signal, which has been long-known in the literature, but due to its need for extensive processing has been scarcely used for diagnostics purposes. After an introduction and overview on the topic and challenges of diagnosing sleep health, we provide the needed mathematical background for the mathematical tools used throughout this dissertation. We then move to the first part of the thesis, where we present algorithms designed to extract information and diagnose sleep using the EDA signal. We developed these algorithms using both traditional signal-processing and ML methods. After presenting data-driven tools, we dedicate the second part of this thesisto develop physics-driven frameworks to study thermoregulation, which greatly affects the EDA signal. During sleep, the EDA signal is generated through a combination of sleep-specific patterns and thermoregulation responses. Furthermore, there is plenty of empirical knowledge on human body thermoregulation and the optimality of such processes during specific part of the night. We start by proposing a novel framework using control theory and mathematical biology to translate experimental observations on physiological mechanisms to a small set of ordinary differential equations (ODEs). As many others processes during sleep, thermoregulation dynamics undergoes significant and abrupt changes making it, approximately, a switched system. Characterising these class of dynamical systems is an open problem in control theory; even proving the stability of such systems in the simplest of cases, for example, linear switched systems, can be a challenge. We conclude this dissertation proposing an overlook on how the methods and ideas introduced in this thesis can be used to formulate and tackle new research problems.