2026 Theses Doctoral

Translating Digital Health to Unconstrained Environments: Behavioral, Biomechanical, and Algorithmic Innovations in Remote Health Monitoring

Chern, Terry L.

Digital health technologies, including mobile health (mHealth) applications and continuous wearable sensors, offer unprecedented opportunities to manage infectious and chronic diseases outside traditional clinical settings. However, transitioning these technologies from controlled environments to reliable real-world utility requires overcoming a triad of challenges: complex behavioral gaps, dynamic biomechanical confounders, and algorithmic representational limits. This dissertation presents three interconnected studies that systematically investigate these requirements across the domains of infectious disease screening and continuous ambulatory blood pressure (BP) monitoring.

First, to address the behavioral "intent-to-action" gap in infectious disease screening, we conducted a 95-participant randomized controlled pilot trial evaluating SMARTest, a custom mHealth application for COVID-19 self-testing, with each participant tracked over a continuous four-week study period. While baseline proactive intent was high across the cohort, multivariable regression analysis revealed that app assignment significantly amplified total testing volume, specifically among highly motivated users. These findings demonstrate the platform’s capacity to serve as a critical activation mechanism, converting pre-existing intent into measurable behavioral adherence within a general population, over a four-week period.

Second, we addressed a ubiquitous biomechanical confounder in continuous wearable BP moni- toring: hydrostatic pressure artifacts induced by free-living movement. Utilizing a 74-participant ambulatory protocol, we developed a multimodal pipeline using minimal inputs, continuous upperarm two-axis accelerometry and baseline participant height, to passively estimate vertical sensor offset from heart level. By decomposing the artifact into central body orientation (CNN-LSTM; AUC 0.81–0.95), static device position (data-driven heuristic; MAE 1.52 cm), and dynamic arm position (Random Forest; 72.2% accuracy), the combined physical model successfully corrected hydrostatic pressure variations to within 1.66–3.67 mmHg, significantly improving the accuracy of unconstrained ambulatory monitoring.

Finally, building upon these measurement requirements, we optimized the algorithmic architectures for cuffless BP classification using an upper-arm multi-wavelength photoplethysmography (mwPPG) device. We evaluated the predictive efficacy of distinct temporal geometries (45-second windows, 5-second sub-windows, and foot-to-foot pulse segmentation) alongside feature representations (learned deep temporal embeddings versus tabular pulse wave analysis [PWA]). Our results demonstrate that generalized, population-level morphology fails in zero-shot transfer. However, the integration of raw deep pulse embeddings with static anthropometric covariates and brief, person-specific calibration fine-tuning yielded much improved ordinal agreement and off-by-one accuracy.

Together, this work advances the paradigm of remote health monitoring. By validating a targeted mHealth behavioral intervention and establishing robust biomechanical and algorithmic frameworks for wearable hemodynamic tracking, this dissertation provides a comprehensive roadmap for deploying reliable, patient-centered digital health solutions in unconstrained, real-world environments.