Researchers have demonstrated that electronic health record foundation models can be trained across multiple healthcare institutions without requiring patient-level records to be pooled in a central database.

The study evaluated federated training of generative event models using 122,251 intensive-care hospitalisations from the University of Chicago Medical Center, Northwestern Medicine and Beth Israel Deaconess Medical Center.

The records were harmonised using the Common Longitudinal ICU Data Format and represented as chronological sequences of clinical events. The models were evaluated across 12 prediction tasks using local, cross-site, centralised and federated training approaches.

Generative event models recorded the highest average performance within individual institutions and when transferred between health systems. Their average cross-site performance loss was lower than that of conventional supervised models based on LightGBM.

Federated optimisation also approached the performance achieved by centralised multi-site training, with most improvements occurring within five to 10 communication rounds.

Under the approach, participating hospitals exchange model updates rather than transferring their underlying patient records. This could allow healthcare organisations to collaborate on clinical AI development while retaining control of protected health information.

The researchers found that multi-site training was most useful when an institution had limited local training data. Its advantage narrowed as the amount of information available within the target hospital increased.

The study concluded that federated training is technically feasible for electronic health record foundation models, although differences in patient populations, documentation practices and clinical systems continue to limit the transferability of models between institutions.