PPFL: A personalized progressive federated learning method for leveraging different healthcare institution-specific features

Abstract

Federated learning (FL) in healthcare allows the collaborative training of models on distributed data sources, while ensuring privacy and leveraging collective knowledge. However, as each institution collects data separately, conventional FL cannot leverage the different features depending on the institution. We proposed a personalized progressive FL (PPFL) approach that leverages client-specific features and evaluated with real-world datasets. We compared the performance of in-hospital mortality prediction between our model and conventional models based on accuracy and area under the receiver operating characteristic (AUROC). PPFL achieved an accuracy of 0.941 and AUROC of 0.948, which were higher than the scores of the local models and FedAvg algorithm. We also observed that PPFL achieved a similar performance for cancer data. We identified client-specific features that can contribute to mortality. PPFL is a personalized federated algorithm for heterogeneously distributed clients that expands the feature space for client-specific vertical feature information.

Keywords: Artificial intelligence; Machine learning.

Graphical abstract

Figure 1

Performance comparison between PPFL and FedAvg, FedProx, FedPer, FedRep, Local (c), and Local (c, s) in terms of AUROC PPFL (c, s) shows the highest AUROC score for predicting in-hospital mortality task. PPFL, personalized progressive federated learning; HICU, heart intensive care unit; MICU, medical intensive care unit; SICU, surgical intensive care unit; NSICU, neurosurgical intensive care unit; AUROC, area under the receiver operating characteristic; FedAvg, federated averaging; FedPer, federated learning with personalization layers; Local, local algorithms; (c), using only common features; (c, s), using both common and specific features.

Figure 2

SHAP values of common and vertical features in predicting in-hospital mortality (A) Heart intensive care unit. (B) Medical intensive care unit. (C) Surgical intensive care unit. (D) Neurosurgical intensive care unit. Client-specific vertical features are highlighted with a black box. BUN, blood urea nitrogen; BP, blood pressure; RBC, red blood cell; HCT, hematocrit.