Calibrated Candidate Pools for Synthetic EHR Release: Coverage, Concentration, and Privacy Risk

Synthetic and mixed electronic health record (EHR) candidate pools support development when real clinical records cannot be shared, but a candidate release should cover the intended population without concentrating probability mass on a few records. We study this question as a fixed-pool audit. Given candidate trajectories from simulators, public demo EHRs, and sequence-model samples, we fit standard simplex-constrained calibration weights and report representation coverage, held-out distributional discrepancy, effective sample size, maximum weight, source lineage, and a membership-inference diagnostic. On a disjoint eICU Demo reference, native source labels leave 26 unmatched feature coordinates after adding Synthea, MIMIC-IV Demo, MIMIC-IV-on-FHIR Demo, and PromptEHR candidates. A deterministic common-ICU representation leaves one unmatched coordinate. In that harmonized space, augmented pools reduce held-out discrepancy range from 0.395-0.410 before calibration to 0.224-0.234 at a fixed operating point while retaining effective-sample-size ratios of about 0.14-0.15. The experiment motivates reporting fidelity together with representation coverage, weight concentration, lineage, and attack-based privacy evidence.