MBMLPMixer: Multiple Branch Multi-Layer Perceptron-Mixer Network for Multi-class Fine Joint Motor Imagery Decoding

Motor imagery brain-computer interfaces (MI-BCIs) face significant challenges such as distinguishing neural patterns of hand, wrist, elbow, and shoulder movements in decoding Fine Joint Motor Imagery (FJMI) tasks due to minimal spatial differences in EEG signals. To address this issue, we proposed the Multiple Branch Multi-Layer Perceptron Mixer Network (MBMLP-Mixer), a novel architecture that integrates multiscale feature extraction and MLP-Mixer units to simultaneously capture and fuse spatial-temporal information across four FJMI tasks. Evaluated on unilateral upper-limb FJMI-EEG data, the model achieved a 4-class classification accuracy of 46.11%, outperforming typical deep learning methods. This study demonstrates the efficacy of combining multiscale embeddings with MLP-Mixer-based feature integration for multi-class FJMI decoding, highlighting the potential of MLP-Mixer networks in enhancing discriminative capability for EEG-based FJMI-BCI applications.