Foundation Model-Guided Synthetic EHR Release: Performance Enhancement with Privacy Preservation.

Abstract

Machine learning on electronic health records (EHR) is constrained by limited data, strict privacy regulations, and distribution shifts across sites and time. Synthetic data offer a potential solution, but existing tabular generators often struggle in the small sample size in EHR regime. We propose MEDPFN, a distribution-aware synthetic EHR generator built on top of the TabPFN tabular foundation model. MEDPFN uses TabPFN as a learned compatibility score over EHR and adds a Mahalanobis regularizer that anchors generated samples to realistic feature statistics, stabilizing generation from small reference cohorts. On six public tabular EHR datasets, we treat the original data only as private context, train standard classifiers solely on MEDPFN-generated samples, and evaluate on held-out real patients. Across datasets and models, MEDPFN-based training matches or exceeds real-data baselines and outperforms several state-of-the-art synthetic generators, enabling high-utility, privacy-friendly EHR data release.