IPROPS - Iterative Prompt Refinement for Optimizing Privacy-Preserving Synthetic Data Generation

The integration of AI in healthcare is heavily impacted by limitations associated with medical data. Data scarcity, strict privacy regulations, and inherent biases affect the development and deployment of AI solutions in medical settings. One common strategy to combat data limitations is by generating synthetic data and utilizing it for training AI models. We introduce IPROPS—a novel framework for synthetic data generation designed specifically to address the inherent limitations of medical data. IPROPS employs advanced prompt optimization techniques that enable automatic refinement of input prompts in prompt-based models and is designed to operate without requiring direct or explicit access to real patient data. Our framework operates as an integrated pipeline with components that function in an iterative manner to generate high-quality, privacy-compliant synthetic data. To validate IPROPS and demonstrate its practical utility, we implemented and evaluated a prototype. Specifically, we applied our framework to generate synthetic German cardiology discharge letters—a complex medical text generation task requiring both clinical accuracy and strict privacy preservation. Results demonstrate the effectiveness of the actor-critic feedback loop and guided mutation strategies in iterating prompts, ultimately producing synthetic data that closely resembles real data. While opportunities for enhancement remain, the IPROPS framework offers substantial benefits, especially in domains where regulatory constraints and data access restrictions present significant barriers to AI advancement.