PSyGenTAB Generates Privacy-Preserving Synthetic Clinical Data

The development of medical AI is frequently hampered by restricted access to high-quality clinical data due to stringent privacy regulations like HIPAA and GDPR, alongside data silos. Synthetic data generation offers a promising avenue, but current methods often struggle to balance privacy with data utility, potentially degrading crucial clinical patterns or risking patient re-identification. This paper introduces PSyGenTAB, a novel privacy-preserving generative framework that addresses this challenge. It reframes synthetic healthcare data generation as a constrained optimization problem, which is then solved using the Augmented Lagrangian Method. Through extensive evaluation across multiple clinical benchmarks, PSyGenTAB demonstrated its ability to maintain essential inter-feature clinical relationships and minority-class diagnostic patterns. Downstream AI models trained on PSyGenTAB's synthetic data achieved performance comparable to those trained on real patient records, while privacy audits confirmed reduced exact record reproduction and strong resilience against membership inference attacks. This framework provides a principled way to foster secure, cross-institutional AI development in healthcare.