Author ORCID Identifier

https://orcid.org/my-orcid?orcid=0000-0002-2489-1130

Date of Award

Fall 10-30-2026

Document Type

Thesis (Ph.D.)

Department or Program

Computer Science

First Advisor

Andrew T Campbell

Abstract

Foundation models offer an impactful paradigm for healthcare by learning generalizable representations from vast amounts of unlabeled data that can efficiently adapt to a wide range of clinical tasks. This thesis presents five studies organized around three pillars crucial to advance foundation models in healthcare. Pillar I: Pre-training: Chapter I introduces PaPaGei, an open foundation model designed for photoplethysmography (PPG) signals. PaPaGei is pre-trained on large-scale continuous optical data using a morphology-aware self-supervised objective, outperforming general time-series models across diverse clinical tasks while remaining parameter-efficient and robust across skin tones. Pillar II: Alignment: To bridge continuous sensor data and language models without relying on text prompting, Chapter II presents Time2Lang, a model-reprogramming framework that maps time-series foundation model features directly into language spaces. Chapter III introduces LENS, a multimodal architecture that pairs a patch-based encoder with a scalable data synthesis pipeline to transform longitudinal mobile sensing into clinically grounded mental health narratives. Pillar III: Clinical Applications: Moving foundation models into authentic psychiatric workflows, Chapter IV investigates automated psychotherapy fidelity estimation on clinical transcripts, showing that structured task decomposition and evidential reasoning minimize systematic scoring bias. Chapter V evaluates patient-grounded suggestions for therapy session preparation in serious mental illness care, showing through clinician evaluation that incorporating patient data enhances personalization and clinical insight. Overall, this work establishes the groundwork to train foundation models for healthcare and integrate them into clinical practice, moving toward systems that close the loop between continuous patient data and personalized care.

Download

Share

COinS