Empirical Evaluations of Personalized Federated Learning on Heterogeneous Electronic Health Records

Yuqing Shang, Qiming Wu, Siqi Li, Di Miao

Keywords: personalized federated learning, meta learning, electronic health records

Abstract: Beyond mobile health devices, federated learning (FL) in healthcare often occurs in cross-silo scenarios, revealing an underexplored area — the comparison of FL, including personalized FL (PFL) models with pre-existing local models. The fact that the majority of existing FL and PFL algorithms were originally designed for cross-device FL settings leaves potential room for improvements in cross-silo scenarios. Our study tests several PFL frameworks on real-world heterogeneous electronic health records, and we also adapt an existing PFL framework, PerFed-Avg, to cross-silo setting by allowing personalized local epochs in clients. Results show that our modified PFL algorithm can benefit cross-silo clinical structured data, and personalizing local epochs contributes to FL model performance.

Submission Number: 114