Federated Learning for ICU Mortality Prediction: Balancing Accuracy and Privacy in a Multi-Hospital Setting

Abstract Intensive care units (ICUs) manage critically ill patients whose clinical outcomes rely on timely and accurate decision-making. Predictive modeling using electronic health records (EHRs) has shown promise in forecasting adverse events such as in-hospital mortality. However, constructing robust and generalizable models typically requires large-scale, multi-institutional datasets. Privacy regulations such as HIPAA and GDPR make centralized data aggregation across hospitals challenging, posing a barrier to collaborative healthcare research. In this study, we explore a privacy-preserving approach for ICU mortality prediction using Federated Learning (FL), which enables distributed model training without sharing sensitive patient data. Leveraging a curated subset of the eICU Collaborative Research Database—a multi-center dataset of over 200,000 ICU admissions—we simulate a real-world federated learning scenario using data from three distinct hospitals comprising 5,000 ICU stays. We design a federated logistic regression framework and evaluate its performance against independently trained local models. Our findings indicate that the federated model consistently outperforms local baselines, achieving improved accuracy and AUC while maintaining strict data privacy. Our work delivers a reproducible and scalable framework for privacy-preserving healthcare AI and demonstrates the practical feasibility of deploying federated models for critical outcome prediction in ICU settings.