Answers are grounded in this paper and cite their sources.
Federated learning (FL) has attracted growing attention in the field of privacy-preserving electroencephalogram (EEG) decoding, as its capability of enabling collaborative learning of a shared global model among multiple clients (e.g., institutions or subjects). Despite the progress, the discrepant EEG data distribution across clients drastically weakens the generalizability of FL model. In this paper, we propose a structure graph-guided personalized federated learning (SG-pFed) framework to address heterogeneous data clients, thereby improving privacy-preserving EEG decoding performance. Specifically, local clients learn their own class-specific prototypes in the EEG representation space, which are then transmitted to the central server to construct both local and global structure graphs. The divergence between local and global structure graphs further serves as a similarity metric, directing the selective aggregation of local models. This process takes into account the variations in EEG data distributions among local clients. Meanwhile, the similarity metric is also utilized to adaptively adjust the degree of personalization to personalize the global model to better fit the distribution of local EEG data. The proposed personalized FL framework can achieve promising EEG decoding performance under various settings of data heterogeneity, such as covariate shift and label shift. Extensive experiments on three public motor-imagery (MI)-based EEG datasets demonstrated the outstanding performance of SG-pFed over the state-of-the-art methods. Visualization results indicate that SG-pFed enables local models to focus more on MI task-related brain regions, offering the underlying mechanisms for performance improvement. The code is available at:https://github.com/ljxcn97/SG-pFed.
No takes yet. Share an insight, caveat, or question.
Hang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: