Adaptive Generative AI Framework for Creating Rare-Disease Synthetic EHRs with Built-In Bias Mitigation and Privacy Guarantees

Author: Rasel Mahmud Jewel, Mohammad Shafiquzzaman Bhuiyan, MD Habibur Rahman, Ahmed Ali Linkon, Tamanna Pervin, Nafis Anjum

Issue: Spring Issue, 2026

Abstract

This comparative research explores the potential applications of generative artificial intelligence (AI) methods in creating synthetic electronic health records (EHRs) for training medical AI models. Currently, the growing concerns about healthcare data scarcity, stringent privacy restrictions, and the need for diverse datasets have led to the emergence of synthetic EHRs as a promising solution. This study examines the most advanced generative models, including generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion-based methods, to determine which can produce the most realistic and privacy-protected datasets. The current study quantifies the utility of synthetic data in training AI models by performing an extensive comparison based on statistical similarity, downstream clinical predictive performance, and privacy leakage. In addition, synthetic EHR effectiveness is assessed using a case study of chronic disease prediction during simulated low-resource conditions. The results indicate that synthetic EHRs can improve access to clinical data while also highlighting significant challenges and providing recommendations for further research.