SynPre-FL Boosts Federated Learning for Clinical Risk Prediction.

This research introduces SynPre-FL, a comprehensive framework designed to enhance privacy-preserving clinical risk prediction using federated learning (FL). The framework tackles several key challenges in FL deployment, including limited data sharing, diverse client datasets, class imbalances, and the scarcity of realistic electronic health record (EHR) benchmarks. SynPre-FL integrates a high-fidelity synthetic EHR data generator with a synthetic-pretrained FL approach to achieve robust predictions under non-IID (non-independently and identically distributed) conditions. The core of SynPre-FL involves a latent autoencoder-diffusion model that creates privacy-preserving synthetic cohorts. These synthetic datasets are then used to "warm-start" federated training, followed by heterogeneity-aware optimization techniques such as class-balanced local objectives, proximal regularization, and adaptive server aggregation. The framework also incorporates post-hoc calibration and federated-safe explainability to ensure reliable and interpretable risk estimates. Experiments confirm that the synthetic generator preserves data structure while protecting privacy, and SynPre-FL consistently improves robustness and scalability across various federated settings, making it a practical solution for distributed clinical prediction.