Dichotomizing illness from cardiovascular and locomotor activity time series
MetadataShow full item record
This thesis addresses the issue of automated evaluation of severity of illness in psychiatric populations. In particular, given that both physiology and locomotor activity have been shown to be modified during mental illness, this work analyses the potential for the use of these measures to assess the discrimination of mental illness using supervised learning algorithms. In particular it examines the discriminatory power of information in heart rate time series and locomotor activity in three ways: 1) using multiple time scales (from minutes to several days), 2) during specific times (as a proxy for context) and 3) using interactions between locomotor and physiological time series. This thesis is comprised of four parts: 1) a review of past work, 2) classification of mental illness using features from quiescent segments of HR, 3) classification of mental illness using features from both heart rate and locomotor activity time series over varying time scales, and 4) evaluation of coupling and interactions between heart rate and activity as features for classifying illness. In Part 1), the body of work upon which this thesis builds is summarized in a review of digital sensors for neuropsychiatric illness. First, the two specific mental illnesses of focus are discussed: schizophrenia and PTSD. Heart rate variability (HRV) and locomotor activity, as well as relevant metrics and features therein are reviewed. The growing literature on digital sensors for monitoring neuropsychiatric illnesses is surveyed, with a focus on passive monitoring and analyses of HR and locomotor activity, feature extraction, and classification or regression of clinically relevant outcomes. In Part 2), features from heart rate time series data are used to train a classifier to distinguish post-traumatic stress disorder from controls subjects. This work explores the hypothesis that data from quiescent (low activity) segments will be more useful for discrimination than data from other segments during the 24-hour recording. This is driven by the knowledge that sleep minimizes exogenous sources of HRV, such as social routine and physical activity. Dysautonomia detectable via alterations in HRV measures such as LF and HF power may thus be amplified during these quiescent segments. Classification is shown to be improved by segmenting the data using low heart rate segments as a proxy for the most restful period of sleep. In Part 3), the work explores the hypothesis that information relevant to pathologically altered physiology and behavior varies with time scale. Features from both heart rate and locomotor activity data recorded over several days are used to train a classifier to distinguish subjects with schizophrenia from healthy controls. The time scale (e.g. window length) of data is varied and found to affect classifier performance, which has a direct relevance to the practical usage of the classifier. In Part 4), the work explores the hypothesis that information between signals is altered in mental illness and relatively less altered in cardiovascular illness, and that this information is useful in a machine learning approach to discriminate patients from controls. Interactions between heart rate and locomotor activity are evaluated using information theoretical approaches, and found to contribute significantly to the classification of schizophrenia over combining univariate approaches, and differently to the classification of mental versus cardiovascular illness. In summary, this thesis demonstrates that physiological data and locomotor activity data, over multiple time scales, independently provide discriminatory power in evaluating psychiatric conditions. Furthermore, the interaction between the two domains (movement influencing physiology and vice versa) provides significant additional discriminatory power.