Personalized Federated Learning for In-Hospital Mortality Prediction of Multi-Center ICU

Federated learning (FL) enables training across distributed healthcare institutions without sharing raw EHR data, but performance degrades under naturally non-IID and unbalanced data. Using an in-hospital ICU mortality prediction task on the real eICU-CRD multi-center dataset while preserving original distribution skew, the paper analyzes why baseline FL degrades and proposes a personalized federated learning (PFL) approach, POLA (Personalized One-shot Local Adaptation). POLA conducts one-shot, two-step training: global FL to obtain a teacher model, followed by local KD-based adaptation with automated personalization to produce high-performing client-specific models. Compared with two PFL baselines, POLA improves prediction performance and significantly reduces communication rounds, with generality to other cross-silo scenarios.