AI-driven wearable biosensors for real-time autonomic function assessment

Thesis Abstract

Autonomic dysfunction is a pervasive yet under-detected contributor to cardiometabolic risk, stress-related illness, and performance decrements, necessitating continuous, real-time monitoring that bridges laboratory precision with real-world applicability. This study addresses the gap between episodic clinical assessments and dynamic autonomic regulation by leveraging AI-driven wearable biosensors to quantify autonomic function in daily life. The aim is to develop a validated, scalable framework for real-time autonomic assessment and to identify patterns predictive of adverse health outcomes. Specific objectives include (1) designing a multimodal wearable sensor platform capable of capturing heart rate variability, photoplethysmography, skin conductance, respiration, and posture; (2) extracting robust physiological features and temporal patterns using machine learning models; (3) evaluating the diagnostic and prognostic value of autonomic indices against established clinical benchmarks; and (4) assessing user adherence, data quality, and interpretability of AI-driven outputs for non-specialist users. The study adopts a mixed-methods, longitudinal design conducted in both controlled laboratory and real-world settings. A total of 320 adult participants aged 25–65, stratified by baseline autonomic tone and risk factors (e.g., hypertension, anxiety, diabetes), will be recruited from two metropolitan communities. Quantitative data will be collected via a validated wearable platform incorporating ECG-derived metrics, PPG-based arterial stiffness, galvanic skin response, respiration rate, and accelerometry, synchronized with pain, sleep, and activity diaries. For qualitative insight, a sub-sample of 40 participants will engage in semi-structured interviews to explore perceived usability and interpretability of AI outputs. Data analysis will proceed in parallel physiologic data will be preprocessed and segmented into 5-minute windows, followed by feature extraction including time-domain and non-linear metrics (e.g., RMSSD, LF/HF ratio, sample entropy). Machine learning pipelines will be implemented using supervised and semi-supervised approaches, including random forest, gradient boosting, and recurrent neural networks, with model selection guided by cross-validated predictive accuracy and clinical relevance. Model explanations will be enhanced via SHAP values to elucidate feature contributions. Autonomic function indices will be benchmarked against gold-standard autonomic tests (Ewing battery) and clinical risk scores. Validity and reliability will be established through test-retest analyses, intraclass correlation coefficients, and Bland-Altman agreement with reference measures. Ethical considerations include informed consent, data privacy, and secure, compliant handling of health information per GDPR and institutional review board guidelines. Anticipated findings indicate that AI-driven wearable biosensors can detect subtle autonomic shifts associated with stress exposure, physical exertion, and sleep fragmentation, with predictive validity for short-term episodes of adverse events such as tachyarrhythmias and hypertensive surges. The integrative model is expected to achieve higher diagnostic accuracy for autonomic imbalance than single-sensor approaches, with AUROC improvements of 0.08–0.12 over baseline metrics. The study also expects to identify diurnal patterns and context-specific autonomic signatures (e.g., postural change, cognitive load) that enhance interpretability for clinical and lay users. The contribution to knowledge encompasses (i) a robust, scalable framework for real-time autonomic assessment in everyday environments, (ii) empirical evidence on the utility of multimodal AI analytics in autonomic physiology, and (iii) practical insights into user-centered design and communication of AI-derived health signals. The study concludes that continuous AI-enabled autonomic monitoring provides actionable insights for early intervention and personalized health management, particularly for populations at high risk of cardiovascular and stress-related disorders. Recommendations include integration with clinical workflows for risk stratification, development of adaptive alert thresholds to minimize alarm fatigue, and further validation in diverse settings and across age groups. Limitations relate to potential data heterogeneity and sensor wearability adherence, which will be mitigated through robust preprocessing, calibration routines, and user training.