Synthetic tabular healthcare data in health informatics: fixed-size substitution and incremental-value controls for predictive utility Application Paper

Abstract Adding synthetic records can make a predictive model appear better simply because it receives more training rows—not because synthetic data add useful information. Repeated predictor profiles split across training and test sets can create a second source of overoptimism. We designed a benchmark that controls both problems. Using three public diabetes-related healthcare datasets, we evaluated Gaussian Copula, Conditional Tabular Generative Adversarial Network (CTGAN), and Tabular Variational Autoencoder (TVAE) with logistic regression, random forest, and histogram gradient boosting across 10 paired replications. First, we held classifier-training size constant while replacing 0% to 100% of real rows with synthetic rows. Second, we compared N real rows (RR-N), N real plus N bootstrap-real rows (RB-2N), and N real plus N synthetic rows (RSADD-2N). Exact duplicate predictor profiles were kept within a single data partition. At fixed size, synthetic substitution produced significantly lower area under the receiver operating characteristic curve (AUC) in 89 of 108 model-specific contrasts after Holm correction, significantly higher AUC in four, and no significant difference in 15. In the additive comparison, synthetic augmentation was significantly worse in 21 of 27 contrasts against RR-N and 16 of 27 against RB-2N. Favourable effects occurred, but were isolated rather than systematic. Under the evaluated conditions, synthetic data therefore offered no general predictive-utility benefit. The study provides a practical framework for separating synthetic-data value from sample-size effects and partition leakage. A reproducibility package supplies code, configurations, public inputs, and validated outputs. These findings do not establish privacy, fairness, clinical equivalence, or deployment suitability.