PERSONALIZED FEDERATED LEARNING FOR IN-HOSPITAL MORTALITY PREDICTION OF MULTI-CENTER ICU

Abstract

Federated learning (FL) presents a promising solution for addressing the challenges of applying machine learning (ML) to privately distributed data, particularly in healthcare settings with multiple independently operated institutions. However, the inherent non-IID (non-identically distributed) and unbalanced nature of data distribution can hamper FL's performance and deter institutions from participating in training. This study investigates these challenges using real-world multi-center ICU electronic health record data, preserving the original non-IID and unbalanced distribution. Initially, the paper examines why baseline FL performs poorly under these conditions before introducing a personalized FL (PFL) approach called POLA to mitigate these issues. POLA is a personalized one-shot, two-step FL method designed to produce high-performance personalized models for each participant. Comparative experiments with two other PFL methods demonstrate that POLA not only enhances prediction accuracy and reduces communication rounds but also exhibits potential for application in similar cross-silo FL scenarios. Its versatility and scalability suggest broader applicability across diverse domains beyond healthcare.