Enhanced Conditional GAN for High-Quality Synthetic Tabular Data Generation in Mobile-Based Cardiovascular Healthcare.

Abstract

Listen

Translate article

The generation of synthetic tabular data has emerged as a critical task in various fields, particularly in healthcare, where data privacy concerns limit the availability of real datasets for research and analysis. This paper presents an enhanced Conditional Generative Adversarial Network (GAN) architecture designed for generating high-quality synthetic tabular data, with a focus on cardiovascular disease datasets that encompass mixed data types and complex feature relationships. The proposed architecture employs specialized sub-networks to process continuous and categorical variables separately, leveraging metadata such as Gaussian Mixture Model (GMM) parameters for continuous attributes and embedding layers for categorical features. By integrating these specialized pathways, the generator produces synthetic samples that closely mimic the statistical properties of the real data. Comprehensive experiments were conducted to compare the proposed architecture with two established models: Conditional Tabular GAN (CTGAN) and Tabular Variational AutoEncoder (TVAE). The evaluation utilized metrics such as the Kolmogorov-Smirnov (KS) test for continuous variables, the Jaccard coefficient for categorical variables, and pairwise correlation analyses. Results indicate that the proposed approach attains a mean KS statistic of 0.3900, demonstrating strong overall performance that outperforms CTGAN (0.4803) and is comparable to TVAE (0.3858). Notably, our approach shows lowest KS statistics for key continuous features, such as total cholesterol (KS = 0.0779), weight (KS = 0.0861), and diastolic blood pressure (KS = 0.0957), indicating its effectiveness in closely replicating real data distributions. Additionally, it achieved a Jaccard coefficient of 1.00 for eight out of eleven categorical variables, effectively preserving categorical distributions. These findings indicate that the proposed architecture captures both distributions and dependencies, providing a robust solution in supporting mobile personalized cardiovascular disease prevention systems.