Domain-Specific Constraint-Aware Tabular Variational AutoEncoder for Synthetic Data Generation

The growth of healthcare artificial intelligence (AI) in resource-constrained regions, particularly in Africa, faces significant challenges including limited access to quality datasets, strict privacy regulations, and inadequate data sharing infrastructures. While synthetic data generation emerges as a promising solution, existing approaches like Tabular Variational AutoEncoder (TVAE) primarily focus on statistical similarity, producing data that, although statistically similar, often lacks domain specific coherence and is unsuitable for critical applications such as medical simulation and clinical training. To address this limitation, we propose Domain-Specific Constraint-Aware TVAE (DSCA-TVAE), an extension of traditional TVAE that incorporates an end-to-end logical constraints loss directly in the loss function. Our approach defines a set of logical constraints (e.g., blood pressure relationships, diabetes correlations, lifestyle consistency) that validate and correct synthetic data generated by a standard TVAE model, ensuring clinical relevance and reliability for medical simulation applications. We evaluated our approach using cardiovascular survey data from northern Senegal (N=1017, 169 features after preprocessing) with 5 independent runs for each model(TVAE, CTGAN and DSCA-TVAE). A constraint compliance score was measure to evaluate the degree of constraint satisfaction, statistical fidelity was assessed using Kolmogorov-Smirnov test for numerical features and chi-square test for categorical features. Machine learning utility was evaluated using the Train-Synthetic-Test-Real (TSTR) framework with Random Forest classifier. Distance based algorithms such as DCR and NNDR was used to assess the privacy preservation of our model compared to two baselines , namely TVAE and CTGAN. DSCA-TVAE achieves an overall aggregated CCR of $0.994\pm 0.003\%$ and fidelity score of $51.56\%$ and AUC value of $91.6\pm 0.0036$ against $97.4\%$ for real data AUC and $93.8\%$ for TVAE($p=0.6370$). For privacy preservation, DSCA-TVAE likely achieves the same results as TVAE and the risk to re-identification of real data is not very concerning. These findings support constraint-aware synthesis as a practical strategy for health AI development in data-constrained settings.