Synthetic Data Benchmark for Small-Sample Parkinson's Disease Rehabilitation Trials

Reproducibility package for a comparative benchmark of tabular synthetic-data generators (Gaussian Copula, CTGAN, TVAE, CopulaGAN) applied to a non-randomised pilot rehabilitation trial in Parkinson's disease (Control n=10, Experimental n=12). It contains the anonymised change-score dataset, a deterministic benchmark script, a null-model Type I error (calibration) script, and the multi-seed supplementary table. Methods are evaluated across five dimensions — distributional fidelity (KSComplement), correlation preservation, clinical plausibility, inferential robustness, and concordance — over 10 fixed random seeds. The analysis is fully reproducible: numpy's RNG is re-seeded with a SeedSequence-derived value before each model fit and each draw (a plain seed does not control the copulas sampler once fitting has advanced the RNG state). Fidelity uses the continuous KS D statistic (KSComplement) rather than a p-value threshold, which is unstable at this sample size.