Adaptive Differential Privacy in Federated Learning for Privacy-Preserving ICU Clinical Decision Support at the Edge

Abstract

Clinical decision assistance that protects patient privacy is a prominent research issue in federated learning, edge computing, and intensive care analytics. However, typical differential privacy algorithms allocate equal privacy budgets to all clinical parameters regardless of sensitivity and predictive power. This may lower privacy-utility trade-offs. This study develops an adaptive differential privacy mechanism for federated learning that dynamically adjusts per-feature Laplace noise and privacy budgets to improve privacy protection while preserving clinical prediction performance. Adaptive Differential Privacy in Federated Learning (ADP-FL) dynamically adjusts per-feature Laplace noise and privacy budgets based on feature-scored sensitivity considering clinical relevance, statistical impact, and re-identification risk. The method was evaluated on the eICU Collaborative Research Database containing 113,265 ICU records and compared against Uniform Differential Privacy and Time-Adaptive Differential Privacy across different privacy budgets (ε). Experiments conducted on the eICU Collaborative Research Database achieved test accuracies of up to 82.7%. Across the evaluated privacy budgets, Feature-Adaptive Differential Privacy consistently outperformed both Uniform and Time-Adaptive Differential Privacy, achieving an F1-score of 0.6672 at ε = 0.5 while maintaining performance close to the no-DP baseline. These findings demonstrate that sensitivity-aware feature-level privacy budgeting preserves key predictive dimensions while minimizing unnecessary perturbation of less sensitive features, providing a robust foundation for privacy-compliant edge-based clinical decision support systems. The novelty of this study lies in its feature-scored adaptive differential privacy mechanism, which dynamically allocates per-feature privacy budgets and Laplace noise in federated learning for privacy-preserving edge-based clinical decision support.