LLM-Generated Expert Summaries Enhance ICU Mortality Prediction

A study published on arXiv (2411.16818) evaluates a multi-representational framework for predicting in-hospital mortality (IHM) in intensive care unit (ICU) patients. The framework fuses large language model (LLM)-generated expert summaries of clinical notes with physiological data. Using the MIMIC-III database, the study analyzed 19,211 first ICU stays with a 12.83% mortality rate. The researchers encoded 48-hour physiology, clinical notes, and LLM summaries (generated under a prompt forbidding prognostication) and then fused these representations. On 3,843 held-out stays, the fusion achieved an AUPRC of 0.4977 and an AUROC of 0.8429, compared to 0.3625 and 0.7770 for physiology alone. Redundancy analysis via ridge regression showed that note embeddings explained 40.8% of the variance in summary embeddings. A note-orthogonal residual retained a minority of the gain (+0.0258 AUPRC), indicating that the summaries provide some non-redundant information. The study concludes that LLM-generated summaries can improve mortality prediction, though much of the benefit overlaps with the notes themselves.