JMIR Publications

ABSTRACT

Background:

The digitization of healthcare has generated vast electronic health records (EHRs) with immense potential for research, yet privacy regulations like HIPAA and GDPR severely limit data sharing. Generative Adversarial Networks (GANs) enable the creation of privacy-preserving synthetic datasets that mimic real distributions. However, standard GANs frequently produce clinically invalid samples, violating biological or logical constraints (e.g., impossible age-diagnosis combinations or invalid physiological correlations), creating a "validity gap" despite strong statistical fidelity.

Objective:

This study aims to bridge the validity gap in GAN-based clinical tabular data synthesis by introducing an inference-time strategy that enforces clinical constraints without retraining models, applying destructive post-processing, or compromising statistical fidelity, privacy, or downstream utility.

Methods:

We trained Wasserstein GAN with Gradient Penalty (WGAN-GP) models, enhanced with spectral normalization and soft constraint penalties, on three public clinical datasets representing varying constraint complexity: Stroke Prediction (N=5,110; simple bounds), Diabetes Health Indicators (N=100,000; intermediate implications), and Cardiovascular Disease (N=68,599; complex multi-variable correlations). At inference, we applied Constraint-Aware Best-of-N Sampling: generating N latent vectors, producing candidate samples via the generator, scoring each for constraint violations (sum across sex-specific, gender encoding, physiological range, correlation, and logical categories), and selecting the lowest-scoring (most valid) sample. Performance was evaluated on validity (violation reduction), fidelity (Kolmogorov-Smirnov statistics, correlations), utility (train-on-synthetic-test-on-real AUC), and privacy (membership inference attack AUC).

Results:

Best-of-N sampling yielded substantial validity improvements with minimal fidelity loss. For simple and intermediate constraints, Best-of-16 achieved 100% violation reduction at ~56ms per sample. For complex constraints (e.g., blood pressure relationships), Best-of-32 reduced violations by 92.4% (Cohen’s d=2.98). A non-linear scaling law emerged, with Best-of-16 as an efficiency sweet spot and a "complexity wall" requiring N=128 for near-perfect validity in structural dependencies. Fidelity remained high (ΔKS<0.03), privacy was preserved (MIA AUC≈0.50), and utility improved (+0.8% to +18.6% TSTR AUC).

Conclusions:

Constraint-Aware Best-of-N Sampling provides a practical, efficient inference-time solution for generating clinically valid synthetic tabular data using GANs. It offers a deployable playbook balancing computational cost and rigor, outperforming soft training penalties while avoiding distributional distortions from post-processing. This approach facilitates safer sharing of synthetic EHRs for medical research and AI development.