SynPre-FL framework boosts federated learning with synthetic EHR data

Federated learning (FL) enables collaborative model training without sharing raw patient data, but clinical deployment is hampered by data scarcity, class imbalance, and client heterogeneity. A new academic paper introduces SynPre-FL, a unified framework that integrates synthetic data generation directly into the FL pipeline. The approach uses a latent autoencoder-diffusion model to produce realistic synthetic EHR cohorts that preserve univariate, bivariate, and multivariate statistical properties. These synthetic cohorts are then used to pretrain a global model before federated fine-tuning, effectively warming up the training process and mitigating cold-start issues. During federated optimization, SynPre-FL applies class-balanced local objectives, proximal regularization, and adaptive server aggregation to handle non-IID data distributions across clients. The framework also includes post-hoc calibration and federated-safe SHAP explainability for reliable, interpretable risk estimates.

Experiments evaluated SynPre-FL across 5, 10, and 15 heterogeneous clients using real-world tabular EHR benchmarks. The synthetic generator proved robust against membership inference and reconstruction attacks, ensuring strong privacy protection. Under TSTR, TRTS, and model-based evaluations, the generated data achieved high downstream utility. In federated settings with severe non-IID fragmentation, SynPre-FL consistently outperformed baseline methods in terms of accuracy, robustness, and scalability. Calibration metrics improved, and SHAP feature attributions remained stable and clinically coherent across different federation sizes. The framework offers a practical, reproducible blueprint for combining synthetic data with federated optimization to enable privacy-aware, interpretable, and robust clinical prediction from distributed EHR data.