Time Series Analytics for Decision Support in Chronic Diseases: Clinical case studies

Chronic diseases burden patients with unending symptoms and functional decline, which limit the activities of daily living and decrease working ability. They increase the risk of injuries, comorbidities, and death. The disabilities from chronic diseases are a major contributor to disease burden globally. With the aging and increasingly obese population, chronic diseases are becoming increasingly common.

Time series analytics offer means to investigate the evolution of chronic diseases over time. The analysis of time dependent patterns can facilitate diverse applications for clinical decision support. Modern analytical methods have grown extremely powerful with the accelerated development of computational resources, being able to mine vast amounts of data and enabling the discovery of all the more complex patterns. Moreover, modern sensor technologies and electronic health record systems have boosted the continuous buildup of high quality health data, monitoring physiological events at hospitals and throughout everyday life. This thesis presents four studies that delve into chronic disease related algorithms across various application time spans, ranging from overnight to several months. The thesis centers around cardiorespiratory measurements collected from healthy and chronic disease patients, measured at hospitals, in free-living settings, and in a controlled laboratory environment.

The studies cover contact-free overnight vital sign monitoring for sleep apnoea detection, wearable sensor based continuous monitoring for fatigue and sleep assessment in neurodegenerative and immune-mediated inflammatory diseases, and sixmonth mortality risk prediction from electronic health records in cardiac patients. The work applies traditional model driven signal processing as well as the more recently emerged data driven deep learning methods, such as transformer neural networks. This thesis presents pragmatic insights on building time series based decision support tools for chronic disease management, and addresses the requirements and limitations related to time series analytics and the underlying data collection across the above-specified time spans. Robust algorithms for contact-free vital sign monitoring are presented and evaluated in broad physiological conditions, the feasibility of continuous monitoring in outpatients and the diverse measurement associations with health related quality of life are analyzed, and the benefits of applying deep learning on health records but also their disadvantages in clinical use are presented. The results imply the importance of high frequency data in applications with short time spans, data collection context tracking in continuous monitoring, and data quality and coverage across all application time spans. The algorithms proposed in this thesis are validated with data collected from human volunteers, including chronic disease patients from the selected disease groups.