Training Large Language Models to Predict Clinical Events

Researchers adapted large language models to predict future clinical events from longitudinal hospital notes.

The paper turns MIMIC-III notes into 6,900 question-answer style examples from 702 admissions using Foresight Learning.

A small LoRA adapter outperformed a prompted base model, improved calibration, and reduced Brier score.

On held-out questions, it slightly beat GPT-5 point estimates.