Publication Date
Spring 2026
Degree Type
Thesis
Degree Name
Master of Science (MS)
Department
Applied Data Science
Advisor
Mohammad Masum; Guannan Liu; Saptarshi Sengupta
Abstract
This thesis presents a unified framework for multi-target wellness prediction using passive wearable sensor data and large language models (LLMs). The study integrates two complementary paradigms: prompt-based regression and representation-driven classification. In the first stage, structured wearable representations derived from daily physiological signals are used to prompt LLMs to predict stress, fatigue, sleep quality, and readiness. Two input representations are evaluated: statistical summaries and hybrid inputs combining statistical summaries with raw sensor values. To improve interpretability and prediction consistency, the second stage introduces a representation-centric classification framework in which wearable signals and wellness targets are discretized into categorical levels. Multiple fine-tuning strategies, including LoRA, full fine-tuning, and API-based adaptation, are evaluated using categorical and hybrid numerical-categorical representations. The proposed classification framework enables simultaneous prediction of all four wellness targets within a unified structured output format. Experimental results demonstrate that representation design plays a critical role in adapting LLMs to wearable time-series data and provides a scalable approach for multi-target wellness prediction using LLMs.
Recommended Citation
Yu, Cheng-Huan, "Interpretable Wellness Prediction from Passive Wearable Sensor Data Using Large Language Models" (2026). Master's Theses. 5756.
DOI: https://doi.org/10.31979/etd.8d6j-fqeh
https://scholarworks.sjsu.edu/etd_theses/5756
