Refine a model on hospital records improves diagnosis, and patient data leak

A Yale study published in Nature Communications shows that fine tuning language models on real files increases their clinical performance but also their tendency to return sensitive information.

Yale researchers, led by Anran Li as first author and Qingyu Chen as lead author, published on September 14, 2026 in Nature Communications an analysis of memorization in language models applied to medicine. The finding is an uncomfortable trade-off: refining a model on real hospital records does improve its diagnostic performance, but at the same time increases the likelihood that it will render sensitive information seen during training.

Two factors stand out as aggravating: the size of the model and the length of the input sequences. The authors also tested the levers usually invoked to limit the phenomenon, such as adjusting the generation temperature, and consider them to have a marginal effect — a sign that memorization is due to the training architecture and the exposure to the data, not to the output settings.

The resulting recommendation is aimed directly at hospital deployments: subject any refined model to a formal confidentiality assessment prior to clinical commissioning, and use training methods designed to limit retention. The topic goes beyond regulatory compliance, as it relates to the possible re-identification of patients from model discharges.