This paper addresses the challenge of applying federated learning (FL) to in-hospital mortality prediction using a multi-center ICU electronic health record (EHR) database. The inherent non-IID and unbalanced nature of EHR data degrades the performance of standard FL. The authors propose a personalized federated learning (PFL) approach called POLA, a personalized one-shot and two-step FL method that generates high-performance personalized models for each participating institution. Experiments demonstrate that POLA effectively improves prediction performance and reduces communication rounds compared to baseline FL and other PFL methods.
