Federated Learning for Privacy-Preserving Severity Classification in Healthcare: A Secure Edge-Aggregated Approach

Federated learning (FL) has emerged as a promising paradigm for privacy-preserving machine learning across decentralized healthcare systems. This study proposes a secure and adaptive FL framework tailored for multi-institutional healthcare environments, combining structured electronic health records (EHR) and real-world ICU datasets (MIMIC-III) to predict patient severity levels. The framework incorporates secure multiparty computation (SMPC) with Shamir’s Secret Sharing to ensure encrypted communication between clients and edge aggregators, preserving data confidentiality throughout the training process. A key enhancement in this work is the integration of a dynamic edge thresholding mechanism that filters client updates based on round-wise gradient variance. Unlike static thresholds, this adaptive strategy enables real-time decision-making to accept or reject updates, improving robustness against noisy or unstable contributions and simulating real-world client dropout. The system was evaluated on both synthetic and MIMIC-III datasets using CatBoost, XGBoost, and TabNet across multiple threshold configurations and client setups. Performance metrics were reported with statistical confidence, standard deviation and 95% confidence intervals across five independent runs per model. The proposed framework demonstrates high classification accuracy, scalability across clients, and improved resilience to data heterogeneity and communication noise. It further incorporates deployment-aware considerations such as latency, update frequency, and dropout tolerance, making it suitable for integration in production healthcare networks. Experimental results highlight that dynamic thresholding not only improves model convergence but also contributes to reliable, fault-tolerant learning under practical constraints.