The increasing digitization of healthcare systems has led to the generation of large-scale ICU patient data across hospitals. However, centralized machine learning approaches pose significant privacy risks, limiting data sharing across institutions. This paper proposes a Privacy-Preserving Federated Learning (FL) framework for ICU patient monitoring and decision support. The system enables multiple hospitals to collaboratively train predictive models without sharing sensitive patient data. The framework integrates Secure Multiparty Computation (SMPC), dynamic edge-based aggregation, and robust machine learning models such as XGBoost, CatBoost, and TabNet. A dynamic thresholding mechanism is introduced to filter unreliable updates and improve model stability. The proposed system is evaluated using structured healthcare datasets and real-world ICU data (MIMIC-III), demonstrating improved accuracy, scalability, and robustness under nonIID conditions. Experimental results show that the framework effectively balances privacy preservation and predictive performance, making it suitable for real-world clinical deployment.
Panchetti et al. (Wed,) studied this question.