FAIR-EHR synthetic electronic health record cohorts

Nine synthetic patient cohorts generated for the FAIR-EHR benchmark, which compares GPT-2-family small language models against GAN, VAE, and diffusion generators for synthetic electronic health record generation. Each cohort is a patient-by-code binary matrix over 1,071 ICD-9 codes derived from MIMIC-III, truncated to 3 digits and 4 for E-codes. The generators are an independent-marginals baseline, EHRDiff, a variational autoencoder, CorGAN, medGAN, and DistilGPT2, GPT2-Small, GPT2-Medium and GPT2-Large fine-tuned for 200 epochs. The record also holds the per-metric replicate vectors behind the paper's results table and a dictionary mapping every column to its ICD-9 code. Real MIMIC-III records are not included. MIMIC-III is available from PhysioNet under credentialed access.