Empowering pervasive healthcare : mobile analytics systems leveraging multimodal data

Abstract

Traditional healthcare often faces challenges of being reactive, inefficient, costly, and constrained by geographical and logistical barriers. To address these issues, Pervasive Healthcare aims to overcome these challenges by seamlessly integrating healthcare solutions into everyday life and providing critical insights into public health. In recent years, the increasing availability of diverse data modalities, such as user-environment contexts and ambient sensor signals, has brought new challenges and opportunities to enhance pervasive healthcare.

In this thesis, we present four mobile analytics systems tailored to different application scenarios, ranging from specific health conditions to general health monitoring, empowering pervasive healthcare leveraging multimodal data with machine learning techniques. For personal health, we introduce ADHDLens and CTlens. ADHDLens utilizes data from commercial wrist-worn activity trackers, along with self-reported and parent-reported measures, to diagnose and monitor Attention-Deficit/Hyperactivity Disorder (ADHD) in adolescents. A 12-week longitudinal study demonstrated the feasibility and effectiveness of using wearable technology for ADHD-related applications, identifying key predictors for diagnosis and medication status classification. CTLens is a context-aware health inference system developed to monitor children's physical and mental health based on their contextual features. Using data collected from children during different periods of the coronavirus infectious disease 2019 (COVID-19) pandemic, CTLens offers insights into the health impacts of school closures on children and identifies associated risk factors. For public health, we propose HPlens and HealthPrism. HPLens is designed to identify high-risk populations for pandemic transmission. Using machine-learning models, it predicts COVID-19 cases based on population features and analyzes case distributions to uncover key characteristics of high-risk areas. Evaluated with real-world COVID-19 data from Hong Kong, HPLens effectively predicts COVID-19 cases based on population features and highlights vulnerable groups by analyzing the importance and influence of various population features on the prediction results, identifying key insights and implications for managing pandemic risks. HealthPrism is an interactive visual analytics system designed to support researchers in exploring the impact of various context and motion features on children's health profiles. It offers multi-level analysis capabilities and supports cross-modality comparison, providing insights into the factors affecting children's health. Its effectiveness and usability were validated through quantitative evaluations, case studies, and expert interviews.

Based on the real-world datasets, this thesis demonstrates the effectiveness and usability of these systems through extensive quantitative and qualitative evaluations, including statistical analysis, performance metrics evaluation, case studies, user studies, and expert interviews.