Privacy-Preserving Federated Learning for ICU Mortality Prediction in Hypotensive Patients: Differentially Private Federated Learning with a Personalization Ablation

The main contributions of this work are the development of a hypotension-focused ICU cohort from MIMIC-IV v3.1, the integration of GRU-D with DP-SGD and Per-FedAvg-style personalization, and a comprehensive comparison with centralized, federated, privacy-preserving, local-only, and SOFA-based approaches. The ablation study shows that personalization provides no statistically significant benefit under non-IID and privacy constraints. Additionally, severity and feature-importance analyses identify Glasgow Coma Scale (GCS) as the strongest predictor of mortality in the cohort.