A Multidimensional Evaluation of Privacy-Preserving Generative Models for Neonatal Clinical Tabular Data: Fidelity, Utility, and Realism Trade-offs

Background The limited availability of neonatal clinical data due to privacy constraints, small sample sizes, and severe class imbalance poses significant challenges for data-driven medical research. Privacy-preserving synthetic data generation has emerged as a promising solution; however, the relative performance of different generative models for neonatal tabular data remains insufficiently understood, particularly with respect to fidelity, utility, and realism trade-offs. Methods This study conducted a comprehensive benchmarking of seven generative models—CTGAN, Variational Autoencoder (VAE), Conditional GAN (CGAN), classical GAN, CopulaGAN, Gaussian Copula, and a transformer-based model—using a neonatal tabular dataset comprising 252 records and mixed numerical–categorical attributes. Each model generated 1,000 synthetic samples. Model performance was evaluated along three dimensions: (i) fidelity, assessed using Kolmogorov–Smirnov statistics, Jensen–Shannon divergence, and correlation analysis; (ii) utility, evaluated via Train on Synthetic, Test on Real (TSTR) and Train on Real, Test on Real (TRTR) schemes using Logistic Regression and Random Forest classifiers; and (iii) realism, assessed through adversarial detection using ROC AUC. Results The results reveal a consistent trade-off among fidelity, utility, and realism across all models. CTGAN demonstrated strong categorical fidelity and achieved the highest predictive utility under the TSTR scheme when combined with Random Forest, but exhibited low realism as its synthetic data were easily distinguished from real samples. In contrast, the transformer-based model achieved the highest realism, with adversarial detection performance closest to random discrimination, while maintaining stable numerical distribution fidelity. Copula-based approaches consistently underperformed across all evaluation dimensions. Conclusions No single generative model optimally satisfies fidelity, utility, and realism simultaneously for neonatal clinical tabular data. CTGAN and transformer-based models emerge as the most promising approaches, each excelling in different dimensions. These findings highlight the importance of multidimensional evaluation and application-driven model selection when deploying privacy-preserving synthetic data in neonatal clinical research.