Personalized federated learning with hierarchical reweighting for multi-center clinical prediction.

Background And Objective: Electronic Health Records (EHRs) depicting patient-related information have been accumulated in a distributed manner and significantly contributed to clinical prediction. Although federated learning can collaborate with multiple medical centers without data sharing, the heterogeneity within multi-center EHRs poses a difficulty to achieve satisfactory predictive performance among different individuals. The objective is to train a personalized model for each client that performs well on local data while simultaneously benefiting from federated training.

Methods: In this paper a Personalized Federated Learning (PFL) method named FedRew is proposed. FedRew adopts the problem setup of model agnostic meta learning and conducts hierarchical reweighting both for local adaptation and global aggregation during federated training. For each independent participant, an alternative minimization scheme is tailored to realize sample reweighting and high-performance personalized clinical prediction models are generated. For the central server, a continuously updated weighting mechanism is adopted for aggregation, to ensure local models capable of mitigating data heterogeneity can exert a higher impact.

Results: Extensive experiments were conducted on the collected eICU-CRD dataset containing EHRs from 10 medical centers. The results validated FedRew achieved superior average performance and mean rank, compared to 2 baseline methods and 6 state-of-the-arts PFL methods. For in-hospital mortality prediction, FedRew achieved an average AUROC of 0.894 and a mean rank of 3.2. For remaining length of stay, FedRew achieved an average RMSE of 1.464 and a mean rank of 2.6.

Conclusion: Our FedRew shows competitive performance across clients for two ICU clinical prediction tasks, demonstrating the potential of FedRew in handling data heterogeneity within multi-center EHRs.