Quantifying Fidelity and Utility in Synthetic Healthcare Data

The scarcity of high-quality clinical datasets and strict privacy regulations remain major barriers to developing robust predictive models in healthcare. Generative models offer a promising solution by producing synthetic data (SD) that replicate real data (RD) characteristics while mitigating privacy risks. This study introduces a structured framework for evaluating SD fidelity and utility using the Pima Indians Diabetes Dataset. Three tabular generative models, Conditional Tabular GAN (CTGAN), CopulaGAN, and Tabular Variational Autoencoder (TVAE), were assessed. Fidelity was quantified using the Pairwise Correlation Distance (PCD) and Wasserstein Distance (WSD), while utility was measured through binary classification performance. Optimized generative models generated high-fidelity SD with predictive accuracy comparable to RD. The results demonstrate that PCD and WSD are reliable indicators of SD quality and effective optimization targets for healthcare machine learning applications.