Electronic Theses and Dissertations Archive

Wearable Physiological AI: Turning Real-World Noise into Robustness in the Age of General-Purpose Foundation Models

Date

2026

Document Type

Dissertation

Degree Name

Doctor of Musical Arts

Department

Computer Science

Committee Chair

Santosh Kumar

Committee Member

Brandon Booth

Committee Member

Christos Papadopoulos

Committee Member

Haomiao Ni

Abstract

Wearable physiological AI transforms signals from smartwatches and fitness trackers into health inferences, including blood pressure estimation, sleep disturbance detection, and stress monitoring. Its true utility, however, lies in operating reliably in everyday settings. Physiological signals collected in the wild present challenges. Daily activities introduce motion artifacts, while differences in device fit, wearing location, context, and physiology alter signal quality. Consequently, models trained on clean, curated datasets often struggle to generalize to field conditions. This dissertation investigates whether foundation models can transform large-scale physiological data into robust, generalizable representations for wearable health monitoring. In natural language processing and computer vision, foundation models have transformed machine-learning solutions for real-world problems. We seek to enable a similar transformation in wearable physiological sensing, focusing on wrist-worn photoplethysmography (PPG), a modality widely available through consumer wearables. To learn physiological patterns without discarding real-world variability, we develop Pulse-PPG, a foundation model tailored to PPG. Its motif-based relative contrastive learning framework identifies repeating short temporal patterns that encode physiological information such as heart-rate variability, blood-pressure changes, and vascular dynamics. A learnable distance function compares and aligns similar motifs within noisy data and organizes positive and negative pairs according to their relative similarities. This enables the model to learn robust, adaptable representations for everyday and controlled settings. Pulse-PPG is pretrained using data from the Mobile Open Observation of Daily Stressors (MOODS) study, a 100-day field study of 122 participants. An AI-triggered sampling pipeline prompted participants after physiologically detected events to report stress, stressors, and location while approximately 55,000 hours of raw PPG were collected. Across 11 downstream tasks from five field, laboratory, and clinical datasets, Pulse-PPG outperformed an existing open-source PPG foundation model trained on clinical data and remained competitive with substantially larger general-purpose time-series models. We further extend this framework to a multi-wavelength PPG foundation model through TONE-Net, a skin-tone-agnostic approach that learns relationships among red, infrared, and green signals while preserving amplitude information for SpO2-related inference on laboratory data. Together, these findings show that PPG foundation models can learn from real-world variability and support wearable applications and multi-wavelength optical tasks with potential clinical relevance.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest/Clarivate.

Notes

Open Access

This document is currently not available here.

Share

COinS