Comparative Analysis of Generative AI Techniques for Addressing the Tabular Data Generation Problem in Medical Records

This research paper explores the use of Generative Artificial Intelligence (GAI) techniques to create synthetic medical datasets that balance data realism and privacy preservation. The sensitivity of medical records poses challenges for data access and sharing, making GAI a promising solution. The study evaluates five key approaches—StyleGAN2, CLIP, T5, ViT, and specialized Tabular GANs—across three critical dimensions: distribution fidelity, attribute correlation preservation, and pattern recognition accuracy. The results reveal the strengths and limitations of each technique in generating realistic medical data. StyleGAN2, CLIP, and T5 excel in all dimensions, making them ideal for various applications requiring high-quality synthetic medical datasets. ViT shows promise but may need fine-tuning for specific use cases, while specialized Tabular GANs demonstrate potential but vary in performance. This comparative analysis provides valuable insights for researchers and practitioners at the intersection of Generative AI and healthcare data generation. It underscores the potential of Generative AI in addressing the tabular data generation challenge in medical records, offering realistic and privacy-conscious alternatives for data utilization and model training.