SATrans-Net: Sparse Attention Transformer for EEG-based motor imagery decoding.

Area of Science:

• Biomedical Signal Processing
• Machine Learning
• Neuroscience

Background:

• Brain-computer interface (BCI) systems decode electroencephalography (EEG) signals for motor imagery (MI) to aid individuals with motor impairments.
• Current deep learning models struggle with the long-sequence nature of EEG-MI data, limiting feature extraction and decoding accuracy.

Purpose of the Study:

• To introduce SATrans-Net, an end-to-end framework designed to model long-range dependencies in EEG-MI signals for improved decoding performance.
• To enhance feature extraction and classification accuracy in BCI systems.

Main Methods:

• Utilized two-dimensional depthwise separable convolution (2D-DSC) for spatiotemporal feature extraction.
• Incorporated a Top-K Sparse Attention (TKSA) mechanism within a Transformer architecture to model long-range dependencies efficiently.
• Employed Grad-CAM for Class Activation Topography (CAT) map generation to visualize spatial attention.

Main Results:

• Achieved high cross-session decoding accuracies: 84.72% (BCI IV-2a), 89.76% (BCI IV-2b), and 96.79% (High-Gamma).
• SATrans-Net outperformed existing methods in decoding accuracy.
• Ablation studies confirmed the significant contribution of the TKSA module.

Conclusions:

• SATrans-Net demonstrates superior decoding accuracy and interpretability for EEG-MI signals.
• The model's ability to capture long-range dependencies offers a promising advancement for BCI technology.
• This work highlights the potential of computational techniques in biomedical signal processing for assistive applications.