Development of Wearable Bio-signal Analytics for Real-time Physiological Monitoring

Thesis Abstract

This study addresses the growing need for continuous, noninvasive physiological monitoring through wearable bio-signal analytics to enhance timely intervention and personalized healthcare. The problem centers on the fragmentation of real-time data streams from heterogeneous wearables, limited interoperability, and the challenge of extracting clinically meaningful, individualized insights from multi-modal signals in everyday environments. The aim is to develop an integrated analytics framework that fuses multi-sensor bio-signals for real-time health state assessment and anomaly detection, enabling proactive decision support for clinicians and end-users. Specific objectives include (1) designing a robust data pipeline to harmonize electrodermal activity, heart rate variability, photoplethysmography, accelerometry, and skin temperature from commercial wearables; (2) devising feature extraction and representation techniques that preserve physiologically meaningful patterns across individuals; (3) implementing real-time machine learning models for activity, stress, sleep, and early anomaly detection, with on-device inference and cloud-backed aggregation; (4) evaluating model performance against gold-standard clinical measures in a pilot population and assessing generalizability across diverse user groups; and (5) assessing user acceptability, battery impact, and data privacy implications to inform deployment. The methodology adopts a mixed-methods approach within a prospective observational framework. A multi-site study will recruit 200 adult participants (inclusive of 100 healthy controls and 100 individuals with controlled chronic conditions such as hypertension or diabetes) over a six-month period. Continuous biometric data will be collected via a standardized set of consumer-grade wearables and a research-grade reference device, resulting in an integrated dataset of at least 2,400 person-days of data. Data collection instruments include synchronized multi-sensor wearable devices, a mobile application for contextual labeling (activity type, perceived stress, sleep hygiene), and a clinician-administered health assessment battery. Data preprocessing will address missingness, sensor drift, and inter-device variability, leveraging signal synchronization and imputation using Kalman filtering. Feature engineering will derive time-domain and frequency-domain metrics from heart rate variability, skin conductance, and peripheral perfusion signals, alongside posture and motion features from accelerometry. The analytical framework will deploy on-device lightweight classifiers for real-time inference and a cloud-based ensemble model employing gradient boosting and recurrent neural networks to capture temporal dependencies. Model validation will utilize cross-validation, bootstrapping, and hold-out testing, with performance metrics including AUROC, sensitivity, specificity, precision, recall, and F1-score. Hypothesis tests will examine whether the integrated model improves detection accuracy for stress and sleep disturbances compared with single-sensor baselines. A subsample (n=60) will undergo qualitative interviews to explore user experience, perceived intrusiveness, and privacy considerations, analyzed via thematic analysis aligned with the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT). Key expected findings include that multi-modal fusion enhances real-time detection of stress episodes by at least 15–20% over single-sensor approaches, that on-device inference reduces latency to under 200 milliseconds for critical alerts, and that sleep quality indices derived from combined signals show strong correlation (r > 0.75) with actigraphy-derived measures. The study anticipates robust generalizability across age groups and both sexes, with acceptable battery usage (<5% daily drain) and manageable data transmission loads. Findings will elaborate how specific features, such as skin conductance variability and nocturnal heart rate dynamics, contribute most to stress and sleep state classification, informing best-practice design for wearable analytics systems. The study contributes to knowledge by advancing an end-to-end, interoperable analytics framework for real-time physiological monitoring, demonstrating the feasibility of accurate, privacy-preserving, multi-sensor inference in everyday settings, and providing empirical evidence on user acceptance and deployment considerations for clinical and consumer health contexts. The conclusions will emphasize the potential of wearable bio-signal analytics to enable proactive health management, suggest standardized evaluation protocols for cross-device interoperability, and offer recommendations for regulatory compliance, data governance, and ethics in continuous health monitoring. Recommendations for future work include expanding the sensor repertoire, exploring federated learning to enhance privacy, and conducting long-term impact assessments on clinical outcomes and healthcare utilization.