New AI framework DISCO-TAB uses reinforcement learning to generate realistic synthetic clinical data while protecting patient privacy

DISCO-TAB combines fine-tuned large language models with a multi-objective discriminator system to create synthetic Electronic Health Records (EHR) that are both statistically valid and clinically accurate

The framework evaluates synthetic data generation at four granularity levels: token, sentence, feature, and row, addressing limitations of prior methods that relied on simple scalar feedback

Incorporates Automated Constraint Discovery and Inverse-Frequency Reward Shaping to preserve complex dependencies and handle severe class imbalances common in healthcare datasets

Addresses a critical challenge in clinical AI development: the scarcity of high-quality, privacy-preserving biomedical data needed to train robust decision support systems