Improving Clinical Validity in Synthetic Electronic Health Record Generation Using Best-of-N Sampling: Comparative Evaluation Study

Description

Masud and Hasan test a simple filter in this 2026 study for synthetic medical records, fake patient data generated for research sharing. The problem they target is clinical validity: a record can look plausible on paper while breaking hard rules, like impossible physiology. Their fix, best-of-N sampling, generates many candidate records and keeps the one violating the fewest rules.

They benchmarked two generators on stroke, diabetes, and cardiovascular datasets of 5,110, 100,000, and 68,599 real records respectively. Success depended entirely on whether the generator ever produced valid records at all. Where the valid share was 12 percent, best-of-128 selection cut violations from 88.2 percent of samples to zero in the cardiovascular tests, matching theory exactly, while a generator with zero valid support stayed hopeless. Their favored pairing hit zero violations on all three datasets with strong fidelity to the real data.

Catalog record

Source: https://europepmc.org/article/MED/PMC13422743