Benchmarking fidelity, utility, realism, and privacy tradeoffs among generative models for synthetic neonatal clinical tabular data
Neonatal clinical data are scarce, severely imbalanced, and privacy-sensitive, and the relative behavior of generative models for synthetic neonatal tabular data remains underexplored. We benchmarked seven generative models, CTGAN, Variational Autoencoder (VAE), Conditional GAN (CGAN), classical GAN, CopulaGAN, Gaussian Copula, and a Transformer-based model on 252 neonatal records comprising eight physiological variables and a 30-category nursing diagnosis. All models were trained on one shared stratified 80/20 train–test split over five random seeds and evaluated for fidelity (Kolmogorov–Smirnov statistic, Jensen–Shannon divergence, and correlation difference), utility (Train on Synthetic Test on Real [TSTR], Train on Real Test on Real [TRTR], and Train on Real Test on Synthetic [TRTS] with four cross-validation-tuned classifiers: logistic regression, random forest, XGBoost, and LightGBM), realism (adversarial detection AUC), and privacy (Distance to Closest Record and a distance-based membership inference attack). Differences between models were tested with Kruskal–Wallis tests, Dunn post-hoc comparisons, and Holm/Bonferroni multiple-testing correction. The Transformer-based model and Gaussian Copula showed the strongest fidelity. The Transformer-based model had the numerically highest TSTR F1-macro under all four classifiers (0.156–0.177), but after correction it did not differ significantly from CTGAN, Gaussian Copula, VAE, or CGAN, and its TSTR performance remained below the real-data TRTR reference for three of the four classifiers. Realism was uniformly poor (detection AUC 0.986–1.000): every synthetic dataset was almost perfectly distinguishable from real records. Five of the seven models generated records closer to training than to held-out data; VAE produced the most near-duplicates (4.6%), and membership inference AUC remained near chance (0.46–0.56). Classical GAN showed severe categorical mode collapse. No model optimized all four dimensions. Because the dataset is small and single-center, five seeds limit statistical power, and neither external validation nor clinical review of synthetic records was performed, these findings are exploratory and dataset-specific; they do not indicate that the synthetic data can replace real data for neonatal predictive modeling or are ready for clinical data sharing or deployment.
Authors
- Achmad Noeman (ORCID: https://orcid.org/0000-0003-4858-5338)
- Tb Ai Munandar (ORCID: https://orcid.org/0000-0002-5726-0431)
- Tyastuti Sri Lestari (ORCID: https://orcid.org/0000-0002-0951-7167)
- Ria Arafiyah
- Alimuddin Alimuddin (ORCID: https://orcid.org/0000-0002-7837-3910)
Institutions
- Sultan Ageng Tirtayasa University (ID)
- State University of Jakarta (ID)
- Universitas Bhayangkara Jakarta Raya (ID)
Publication Details
- Journal
- Discover Artificial Intelligence
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1007/s44163-026-02764-1
- Primary Topic
- Privacy-Preserving Technologies in Data
- Type
- article
- Field-Weighted Citation Impact
- 0.00