SSA-DCNet: a cross-session MI-EEG classification network based on deformable convolution and spatial-shift attention

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) using motor imagery (MI) electroencephalogram (EEG) are promising for neurorehabilitation.
• Cross-session variability in EEG signals poses a significant challenge for reliable classification.

Purpose of the Study:

• To develop a robust method for cross-session MI-EEG classification.
• To enhance the performance of BCIs in neurorehabilitation by addressing session-dependent signal variations.

Main Methods:

• Proposed Spatial-Shift Attention Deformable Convolution Network (SSA-DCNet), a compact CNN.
• Utilized 2D deformable convolution for adaptive temporal filtering.
• Implemented a spatial-shift attention mechanism to emphasize stable spatial patterns across sessions.

Main Results:

• Achieved 84.72% accuracy on BCI Competition IV-2a and 90.45% on IV-2b.
• Demonstrated superior discriminative power and robust cross-session generalization via t-SNE visualizations.
• Successfully suppressed session-dependent noise and variability in EEG signals.

Conclusions:

• SSA-DCNet offers a significant advancement in cross-session MI-EEG classification.
• The proposed method enhances the robustness and generalizability of BCIs for neurorehabilitation applications.
• The network effectively captures invariant neural patterns while mitigating session-specific interference.