Training Large Language Models to Predict Clinical Events

TL;DR

This paper introduces a method to train large language models on longitudinal clinical notes for predicting patient events, improving calibration and accuracy without hand-engineered features.

Contribution

It extends Foresight Learning to clinical prediction, creating a scalable approach using natural language questions and labels from longitudinal notes.

Findings

Improved calibration error from 0.1269 to 0.0398

Reduced Brier score from 0.199 to 0.145

Outperformed GPT-5 point estimates on held-out questions

Abstract

Longitudinal clinical notes contain rich evidence of how patients evolve over time, but converting this signal into training supervision for clinical prediction remains challenging. We extend Foresight Learning to clinical prediction by converting time-ordered MIMIC-III notes into examples consisting of past patient context, a natural-language question about a possible future event, and a label resolved from later documentation. This process yields 6,900 prediction examples from 702 admissions across medications, procedures, organ support, microbiology, and mortality. A small LoRA adapter trained on these examples improves over the prompted base model, reducing expected calibration error from 0.1269 to 0.0398 and Brier score from 0.199 to 0.145, while slightly outperforming GPT-5 point estimates on held-out questions. The approach enables reusable clinical prediction supervision from longitudinal notes without hand-engineered structured features or endpoint-specific classifiers.