A health-records AI model holds up in hospitals it was never trained on
Researchers pretrained a model on records from four health systems and tested it on two others, measuring what screening would really look like.
Picture made for ABUREAU.
In a preprint posted on 29 October 2026, researchers describe TRACER, a model that predicts future diagnoses from patients' electronic health records. It was pretrained on records from over 4.5 million patients in four health systems, then tested on 16 diseases in more than 2 million patients from two health systems kept out of training.
The test was built for screening rather than headline scores. At a setting that keeps false alarms low, patients the model flagged were up to eight times more likely to be diagnosed within a year than patients picked at random. The authors also found that detection and false alerts varied by disease and by health system, even where the overall scores looked steady.
The lesson for health data is about breadth. A model is only as useful as the range of records it learns from and is checked against. Records from many separate systems, de-identified and used with proper permission, are what make that possible, and what make each system's own records worth more.