Smartphone-Enhanced Wearable for Real-Time Physiological Stress Monitoring

Thesis Abstract

This study addresses the growing need for unobtrusive, real-time assessment of physiological stress in daily life and occupational settings, where conventional laboratory-based measures fail to capture dynamic fluctuations and context-specific responses. The aim is to develop and validate a smartphone-enhanced wearable system that integrates physiological sensors with a user-facing mobile application to monitor, classify, and interpret stress in real time. Specific objectives include (1) engineering a multimodal wearable prototype capable of synchronizing heart rate variability (HRV), skin conductance, and peripheral temperature with smartphone-inferred behavioral markers; (2) devising an adaptive signal processing pipeline and feature extraction framework to improve signal quality in free-living conditions; (3) constructing a machine learning model to classify stress states across diverse contexts, and (4) evaluating usability, acceptability, and ecological validity of the system in real-world environments. A mixed-methods, multi-site study design is employed with a sample of 320 adults aged 18–65 across three occupational groups (healthcare, finance, education). Quantitative data are collected via the wearable sensor suite (optical HRV, galvanic skin response, skin temperature) and the smartphone app, supplemented by ecological momentary assessment (EMA) prompts (five prompts per day for 14 days) and a baseline laboratory task protocol including the Trier Social Stress Test. Qualitative insight is gained through semi-structured interviews with a purposive subsample of 40 participants to explore perceived stress, usability, and contextual factors influencing data interpretation. Data collection instruments include validated physiological sensors with Bluetooth Low Energy data streams, the STAI-6 for trait anxiety, the Perceived Stress Scale (PSS-10), and the System Usability Scale (SUS). Data analysis employs a robust sequential approach signal preprocessing and feature extraction (time-domain HRV metrics, frequency-domain HRV, skin conductance response features, and peripheral temperature derivatives), followed by supervised learning using nested cross-validation to optimize model generalization. Regression analysis and ANOVA test the association between physiological markers and EMA-reported stress, while a random forest classifier and a gradient boosting model are compared for stress state prediction. Model interpretability is enhanced via SHAP value analysis and feature importance ranking. A Bayesian hierarchical model accounts for within-person and between-person variance in stress reporting. Qualitative data are analyzed thematically using a framework aligned with the Technology Acceptance Model (TAM) and the Job Demands-Resources (JD-R) theory, enabling integration of user experience with physiological indicators. Expected findings include that a multimodal feature set from HRV, skin conductance, and temperature improves stress classification accuracy to at least 85% in daily life, outperforming unimodal baselines. Real-time stress scores demonstrate significant concordance (p < 0.001) with EMA reports, with moderate multi-context variability captured by the hierarchical model. The framework is anticipated to reveal context-specific patterns (e.g., high cognitive load vs. social-evaluative threat) and demonstrate that user engagement and perceived usefulness mediate adherence to data collection. The study contributes to knowledge by (1) advancing an integrative model for real-time physiological stress monitoring in non-clinical populations, (2) validating a scalable, consumer-grade platform linking wearable biosignals with mobile analytics, and (3) elucidating the interplay between physiological signals, contextual factors, and user acceptance in stress management. The main conclusion is that smartphone-enhanced wearables can provide reliable, ecologically valid stress monitoring with actionable insights for individuals and organizations, enabling timely interventions and personalized stress management strategies. Recommendations include refining the calibration protocol to individual baselines, incorporating context-aware prompts to reduce reporting burden, promoting data privacy and user control features, and exploring integration with workplace wellness programs. Limitations include potential sensor motion artifacts, demographic representation, and the need for longitudinal validation beyond the 14-day observation window. Future research could extend the model to clinical populations, examine longitudinal health outcomes associated with real-time stress feedback, and evaluate efficacy in reducing stress-related occupational incidents.