SynPre-FL: Synthetic Data-Driven Pretraining Integrated Federated Learning Training Framework

The paper introduces SynPre-FL, a framework that uses synthetic electronic health records generated by a latent autoencoder-diffusion model to warm-start federated learning. It then applies class-balanced local objectives, proximal regularization, adaptive server aggregation, post-hoc calibration, and federated-safe explainability. According to the abstract, experiments with 5, 10, and 15 heterogeneous clients show improved robustness and scalability, particularly under severe non-IID fragmentation. The authors also report preservation of statistical structure, resistance to membership-inference and reconstruction attacks, strong TSTR and TRTS utility, better probability calibration, and stable SHAP attributions.