FedWFC: Federated learning with weighted fuzzy clustering for handling heterogeneous data in MIoT networks
Le Sun, Shunqi Liu, Ghulam Muhammad · Alexandria Engineering Journal · 2024
The diversity of sources and uneven distribution of medical data contributes to the statistical heterogeneity within the Medical Internet of Things (MIoT) networks. In this context, comprehensive analysis of patient data is imperative to provide more precise diagnoses and treatment strategies, rendering personalized medical treatment indispensable. Moreover, the transmission of medical data over networks raises concerns regarding data privacy, necessitating thorough consideration. To address these challenges, we propose FedWFC, a federated learning method that combines a novel importance weight with fuzzy k-means clustering to effectively handle the heterogeneous medical data in MIoT networks. Firstly, we utilize fuzzy k-means for clustering and partitioning local model parameters from MIoT devices, enabling centralized updates of multiple global models based on these clusters. This cluster-centric approach streamlines personalized updates for local models. Secondly, the introduction of the new importance weight allows us to tighten the optimization error bound for global model updates. Experiments show that FedWFC improves the macro F1 score by 4.24% and the micro accuracy by 4.99% compared with existing methods, highlighting its effectiveness in MIoT data processing. • Federated learning method with fuzzy k-means clustering for data classification. • Distance-constrained local model update for alignment with global updates. • Enhanced classification performance via cluster-centered model updates. • Novel importance weight and regularization term to mitigate incorrect model updates.