Abstract
The burden of healthcare systems is increasing worldwide mainly due to the higher prevalence of chronic disease, aging societies, greater demand for quality healthcare, and rising shortage of healthcare labors. Digital health defines the usage of digital technologies, communication tools, information, data to collect, share, analyze health information. They are promising to achieve self-monitoring, reduce healthcare costs, improve patient or population health and deliver better health care to more people with quality.Health-related data are collected and digitalized tremendously in many ways such as wearable devices, electronic health devices, biomedical sensors, mobile phones, and social networks. To uncover patterns underneath data and extract useful information for monitoring and decision-support, data-driven machine learning (ML) techniques became an essential part of digital health applications.
ML-based approaches can be improved over time and have the potential to generate complementary innovations. With the advent of advanced ML techniques, the challenges and opportunities in digital health are more to leverage it to meet the clinical or medical relevant needs and increase healthcare’s reach, efficiency, and accuracy. A successful ML-based health system requires understanding clinical background, integrating large amounts of heterogeneous information, and selecting appropriate ML models.
In this thesis, we develop needs-oriented and customized solutions based on ML to two major problems: (1) health monitoring using wearable devices and (2) decision support using EHR.
For health monitoring using wearable devices, there are still gaps to map physiological signals to health indicators and pattern changes. Whilst the physiological signals can be collected continuously, health indicators or pattern changes are hard to acquire. We first develop an unsupervised and personalized sleep monitoring algorithm that can achieve sleep/wake identification and sleep pattern changes characterization. Besides, we also propose the integration of heart rate and activity data for model development, and the fusion effect is illustrated with a real-world case study. Secondly, we present a hierarchical blood pressure (BP) estimation paradigm using a photoplethysmography (PPG) sensor to reach care into cerebrovascular disordered patients and to improve the accuracy of estimation. Both proposed ML-based approaches can achieve continuous monitoring and provide accurate and reliable care to more people either through a personalized algorithm or a paradigm adapted to specific phenotypes.
For decision support using EHR, we focus on the incident reports system in hospitals. With the wide adoption of an electronic incident reporting system, manual review becomes impossible when the volume of reports proliferates. We present an incident report classification (IRC) framework to classify in-hospital fall incident severity. Incorporating structural clinical features into traditional IRC frameworks using narrative text only provides informative and essential information to categorize fall incident severity. Resampling methods also help rebalance the class distribution of the original incident report data and improve the classification performance. These findings help move one step closer to automated IRC to standardize the process and minimize the labor.
This thesis contributes to the research on digital health by developing several novel ML-based approaches for different applications in the healthcare field. This thesis also demonstrates that the improvement of digital health applications relies on heterogeneous data fusion, domain knowledge or techniques integration, ML techniques selection, and rigid validation. The novel applications and implementations presented in this thesis provide valuable recommendations and lessons to the field and contribute to the health industry.
| Date of Award | 22 Jan 2021 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Kwok Leung TSUI (Co-supervisor) & Xinyue LI (Supervisor) |