FedSp: Self-Paced Personalized Federated Learning for Clinical Prediction of Multi-Center ICU Patients

Abstract

The development of medical AI faces two key challenges: 1) difficulty of sharing data between medical centers due to privacy and security concerns, and 2) weakness of unified machine learning models in application due to unsatisfactory personalization. Mainstream Personalized Federated Learning (PFL) applies global federated training and local adaptation, but the performance is limited when heterogeneous data distribution exists among individuals, particularly for patients in multi-center ICU scenario. This paper proposes a self-paced PFL framework, FedSp, for clinical prediction. First, FedSp realizes model parameters decoupling to conveniently adopt fine-tuning of the global model with local data. Second, through self-paced learning, FedSp mitigates the adverse effect from heterogeneous data distribution in global learning. The obtained personalized models are experimentally evaluated on a public multi-center ICU dataset, compared with representative PFL methods regarding in-hospital mortality prediction task and remaining length of stay task. The results on 8 medical centers demonstrate the superior performance of FedSp.

Ähnliche Arbeiten