BCIwiki (bciwiki.com) — A new study published in Computing and Informatics on July 14, 2026, proposes a federated graph learning framework named FSDFGL for motor intention recognition in hybrid brain-computer interfaces. The research was conducted by scientists from Shanghai Sunshine Rehabilitation Center and Shanghai Ocean University.
The study highlights that deep learning-based EEG decoding faces multiple challenges: EEG has low spatial resolution, making it difficult to accurately identify specific brain regions; deep learning requires large amounts of data, while EEG datasets are typically small; and merging small datasets raises privacy concerns because EEG signals contain a wealth of personal information. Federated learning (FL) offers a potential solution for limited data and privacy, but it faces the challenge of non-independent and identically distributed (non-IID) data. To address these issues, the team proposed incorporating EMG data into EEG data through SPMI during graph construction, leveraging the distinct features of muscle activity provided by EMG signals. In the local graph neural network (GNN) process, structural information is decoupled from node features, and a structure encoder captures and shares this information globally during FL. The framework was evaluated on the TAN dataset, and experimental results demonstrate its superiority in identifying complex motor intentions.
By sharing structural information rather than heterogeneous features, the framework avoids degradation of local model performance, potentially offering a new approach to data privacy and model generalization in hybrid BCIs.