Paper accepted in JMIR AI 🧬

Clinical Validity in Synthetic EHR Generation: Best-of-N Sampling in Tabular GANs is now published in JMIR AI.

Synthetic EHRs from tabular GANs can look statistically plausible while still being clinically invalid. We formalize this with a three-regime taxonomy based on the generator’s valid-support mass (p_valid), and show that Best-of-N sampling — generating multiple candidates and keeping the one closest to the valid clinical manifold — shifts which regime you land in. Benchmarked across WGAN-GP and CTGAN on multiple clinical datasets, with a privacy audit (distance-to-closest-record, proximity concentration) confirming the gains aren’t just memorization in disguise.