MSTDualNet: multi-scale state-space dual-branch network for electroencephalography-based motor imagery decoding

Abstract

Deep learning has substantially advanced electroencephalography (EEG)-based motor imagery (MI) decoding. However, existing MI-EEG decoding models insufficiently capture temporal dynamics at multiple resolutions, which are important for EEG signals containing transient event-related desynchronization/synchronization (ERD/ERS) changes and sustained rhythmic modulations. To address this issue, we propose MSTDualNet, a dual-branch spatiotemporal framework for within-subject binary MI-EEG decoding. MSTDualNet uses a pyramid multi-scale state-space temporal encoder to model temporal dependencies at different resolutions. A channel-wise spatial Transformer branch captures inter-channel relationships, while a compact multi-level decision module combines prototype-guided classification with an auxiliary temporal convolution branch to enhance feature discriminability. Experiments were conducted on BNCI2014001, BNCI2014002, and BNCI2015001 under the chronological order (CO) and five-fold cross-validation (CV) protocols. MSTDualNet achieved accuracies of 83.81%, 82.07%, and 84.82% under the CO protocol and 87.39%, 83.16%, and 86.12% under the CV protocol, respectively, yielding the highest mean accuracy in all evaluated settings. Compared with DBConformer, a strong structurally related baseline, MSTDualNet improved the average accuracy by 3.02 and 1.80 percentage points under the CO and CV protocols, respectively, while maintaining moderate model complexity. These results highlight the effectiveness of MSTDualNet for within-subject binary MI-EEG decoding and its potential as a multi-scale state-space framework for MI-EEG decoding.