The limited accessibility of healthcare data due to privacy and data-sharing constraints poses a significant challenge for the development of reliable machine learning–based risk prediction models. In this study, the statistical fidelity and machine learning utility of synthetic tabular data generated by the Conditional Tabular Generative Adversarial Network (CTGAN) are systematically evaluated. The analysis is conducted on three datasets with different feature structures: numerical (Hepatitis C), categorical (Diabetes), and mixed-type (cardiovascular disease). The evaluation framework includes three experimental scenarios: Train on Real Test on Real (TRTR), Train on Synthetic Test on Real (TSTR), and Train on Synthetic Test on Synthetic (TSTS). A diverse set of machine learning models is employed, including Logistic Regression, K-Nearest Neighbors, Support Vector Machines, Random Forest, XGBoost, and LightGBM. Model performance is assessed with particular emphasis on metrics sensitive to class imbalance. The results show that the behavior of CTGAN is strongly dependent on the underlying feature structure of the data. In numerical datasets, macroscopic distributions are largely preserved, while notable degradation is observed in representing decision boundaries of rare classes. In categorical datasets, the model shows difficulty in capturing complex multivariate joint distributions, leading to substantial performance drops, particularly for tree-based methods. In mixed-feature datasets, synthetic data appears to introduce a denoising effect that partially stabilizes learning, but at the cost of reduced boundary precision.

Keywords

Risk Assessment, Artificial Intelligence (AI), Healthcare Application, Machine Learning Applications, Data Analytics