Enhancing Clinical Prediction via Synthetic Tabular Data Generation Using Deep Generative Models

Abstract

The development of reliable machine learning (ML) models for clinical diagnosis is hindered by limited availability of large-scale, diverse, and high-quality medical datasets due to privacy constraints, class imbalance, and data heterogeneity. Synthetic data generation using deep generative models provides a promising approach to address these challenges; however, their effectiveness on real-world clinical tabular datasets remains insufficiently investigated. This study presents a comprehensive framework for enhanced clinical classification through synthetic data augmentation using five advanced tabular generative models, including CTGAN, TGAN, COPULA-GAN, CTAB-GAN, and TVAE. The models were evaluated on six heterogeneous clinical datasets with diverse feature characteristics and class distributions. Synthetic data quality was assessed using Kolmogorov-Smirnov and Chi-square statistical tests, while the impact of augmentation was evaluated using Random Forest, Logistic Regression, and Multilayer Perceptron classifiers based on F1 score under 30% and 50% augmentation scenarios. Experimental results demonstrate that TVAE and CTAB-GAN achieved superior synthetic data quality by preserving statistical distributions and feature relationships. Furthermore, synthetic data augmentation improved clinical classification performance by up to 5.2% in F1 score, with optimal performance observed at 300 training epochs and 50% augmentation. The findings demonstrate that high-quality synthetic data generated by TVAE and CTAB-GAN can effectively enhance clinical prediction performance, alleviate class imbalance, and reduce dependency on sensitive patient records. This study provides practical insights for selecting suitable generative models for privacy-aware clinical ML applications and highlights future opportunities in federated synthetic data generation and differential privacy-based frameworks.