FedCVDSyn: Federated Generation of Synthetic Tabular Healthcare Data for Privacy-Preserving Cardiovascular Research

Abstract

Cardiovascular disease (CVD) is a global leading cause of death, but challenges exist in developing strong predictive models due to stringent privacy guidelines and fragmented healthcare data systems. Federated learning and synthetic data production provide two promising options for privacy-preserving solutions; however, most current implementations focus on such distinct aims as federated optimization, differential privacy, or the utility of synthetic data without guaranteeing clinical plausibility or adversarial privacy auditing. To address these issues, it proposes FedCVDSyn, an integrated, multi-objective methodology that generates clinically plausible, privacy-preserving synthetic cardiovascular data across simulated federated clients. The proposed methodology employs Conditional Tabular Generative Adversarial Networks (CT-GANs) that learn collaboratively using Federated Averaging and adaptive differential privacy budgeting without having to transmit actual patient data to other sites. Additionally, source data is screened for clinical rule-compliance using rules derived from the ACC/AHA and ADA guidelines, and Rényi differential privacy accounting is used as an approximate privacy accounting method. Experimental evaluation against four simulated clients of varying data volume demonstrates high statistical fidelity (Wasserstein Distance of 0.0362) and nearly random membership inference attack performance (AUC $= 0.4945 \pm 0.0231$ ). The synthesized data has also achieved a clinical plausibility rating of 0.9993, indicating the ability of FedCVDSyn to produce secure, clinically plausible cardiovascular data.