Evaluating and Enhancing General-Purpose Language Models for Clinical Prediction: Insights from Foundation Models and EHR Benchmarks
Name
shi-lrshi-meng-eecs-2026-thesis.pdf
Size
1.19 MB
Format
Adobe PDF
Checksum (MD5)
3b0effc97347246685d71459d511c7bf
Author(s)
Shi, Lawrence
Advisor(s)
Sontag, David
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Large language models (LLMs) and foundation models (FMs) offer promising frameworks for clinical prediction from structured longitudinal health data, but their effective use requires careful design of patient representations, learning objectives, and calibration strategies. This thesis investigates how reasoning-augmented learning and structured input transformations improve both predictive accuracy and learning efficiency in LLM-based clinical prediction systems across multiple tasks derived from EHRSHOT and claims-based datasets, comparing zero-shot inference, direct supervised prediction, and reasoning-augmented training paradigms. We show that reasoning-augmented approaches enable LLMs to arrive at more accurate clinical predictions by explicitly modeling intermediate decision processes and learning from generated reasoning traces. We demonstrate that structured transformations of patient histories into task-specific clinical schemas further amplify these benefits, leading to more efficient learning and stronger downstream performance. Across a wide range of experiments, the combination of reasoning-augmented supervision and carefully designed representations consistently outperforms the best existing benchmarks in the literature. This demonstrates that reasoning is a core driver of both improved learning efficiency and predictive performance, and that jointly designing reasoning mechanisms and structured representations enables the development of more effective and interpretable machine learning systems for clinical prediction.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
