AI-Driven Mental Health Monitoring via Multimodal Wearable Data Integration

Thesis Abstract

The study addresses the growing gap in scalable, objective mental health assessment by integrating multimodal wearable data with AI-driven analytics to monitor psychological well-being in real-world settings. The problem centers on limited ecological validity of traditional self-report measures and static clinical assessments, which fail to capture dynamic mood fluctuations and context-specific risk indicators. The aim is to develop and validate an end-to-end AI framework that fusion-multimodal physiological, behavioral, and contextual data from consumer wearables to detect early signs of anxiety, depression, and stress-related conditions, while ensuring interpretability for clinical deployment. Specific objectives include (1) to design a data fusion pipeline that harmonizes electrodermal activity (EDA), heart rate variability (HRV), skin temperature, accelerometry, sleep metrics, and smartphone-based contextual cues; (2) to develop machine learning models capable of predicting clinically relevant symptom trajectories over a 12-week monitoring period; (3) to evaluate model performance against standardized instruments such as the Patient Health Questionnaire-9 (PHQ-9) and the Generalized Anxiety Disorder 7-item scale (GAD-7); (4) to assess the interpretability of predictions through SHAP values and clinically oriented rule explanations; (5) to examine ethical, privacy, and user-acceptability considerations in a real-world deployment. The methodology adopts a mixed-methods research design. The population comprises 300 adults aged 18–65 recruited from university-affiliated workers and community volunteers, screened to include individuals with varying baseline levels of depressive and anxiety symptoms. A stratified random sample ensures representation across age, gender, and digital literacy. Data collection employs continuous multimodal sensing from wrist-worn devices (EDA, HRV, skin temperature, accelerometry, sleep stages) and smartphone sensors (ambient context, screen time, geolocation-derived activity). Participants complete weekly PHQ-9 and GAD-7 assessments and a biweekly semi-structured interview to explore user experiences and perceived accuracy of AI feedback. Instruments include validated wearable SDKs, the PHQ-9 and GAD-7 questionnaires, and a bespoke daily diary app to capture subjective affect and life events. Reliability and validity are established through test-retest procedures for self-report measures and calibration of sensor-derived features against clinical benchmarks. Data analysis follows a multi-tiered approach. Data preprocessing includes imputation of missingness via multivariate imputation by chained equations and signal-level denoising. Fusion of multimodal data is achieved using a hierarchical attention-based deep learning model capable of handling temporal sequences and heterogeneous feature types. Model training employs nested cross-validation, with performance metrics comprising AUC-ROC, F1-score, precision, recall, and mean absolute error for symptom trajectory predictions. Interpretability is enhanced through SHAP analysis and clinically oriented rule extraction to translate model decisions into actionable indicators. Model specification includes a time-to-event framework for escalation risk and a multilevel mixed-effects model to examine within-subject variability over time. Hypotheses posit that (H1) multimodal integration yields superior predictive accuracy for symptom trajectories compared to unimodal baselines; (H2) wearable-derived features such as HRV and sleep fragmentation will demonstrate significant associations with PHQ-9 and GAD-7 scores; (H3) model explanations will align with clinical reasoning, fostering higher clinician and user trust. Anticipated findings indicate that the integrated AI framework can detect subclinical shifts in mood and anxiety with clinically meaningful lead times (1–2 weeks) and robust generalization across subgroups. The study contributes to knowledge by advancing a scalable, interpretable, and privacy-conscious framework for continuous mental health monitoring that bridges psychosocial theory and real-time physiological indicators, informed by theoretical underpinnings from the Biopsychosocial Model and the Allostatic Load framework. The implications include improved early intervention, personalized feedback, and evidence-informed guidelines for ethical deployment of AI in digital mental health. The anticipated conclusion emphasizes the feasibility and utility of immersive multimodal wearable data integration for proactive mental health management, with recommendations for integrating clinician oversight, data governance, and user-centered design to maximize acceptability and clinical adoption.