MSHANet: A Multiscale Hybrid Attention Network for Motor Imagery EEG Decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interface (BCI) technology is crucial for neurorehabilitation and motor function restoration in patients with stroke or spinal cord injury.
• Motor imagery electroencephalography (MI-EEG) is a key BCI modality, but its decoding accuracy is hindered by nonlinear dynamics and inter-subject variability.
• Existing BCI methods struggle to effectively capture complex spatiotemporal EEG patterns.

Purpose of the Study:

• To propose a novel Multiscale Hybrid Attention Network (MSHANet) for enhanced MI-EEG decoding.
• To improve the accuracy and robustness of BCI systems for neurorehabilitation applications.
• To introduce an electrode spatial structure-aware encoder to leverage electrode positioning information.

Main Methods:

• Developed MSHANet, integrating spatiotemporal feature extraction (STFE), talking head self-attention (THSA), dynamic squeeze-and-excitation attention (DSEA), and a temporal convolutional network (TCN).
• Incorporated an electrode spatial structure-aware encoder to process electrode positional data.
• Evaluated MSHANet on BCI Competition IV Datasets 2a and 2b, and EEGMMID using within-subject and cross-subject experimental designs.

Main Results:

• MSHANet achieved high decoding accuracies: 83.56% (within-subject, BCI-2a), 89.75% (within-subject, BCI-2b), and 75.66% (within-subject, EEGMMID).
• Cross-subject experiments yielded accuracies of 69.93% (BCI-2a), 81.85% (BCI-2b), and 79.67% (EEGMMID).
• The electrode spatial structure-aware encoder improved decoding performance by up to 2.91%.

Conclusions:

• MSHANet demonstrates superior performance in MI-EEG decoding, outperforming existing methods.
• The proposed network holds significant potential for clinical applications in neurorehabilitation and motor function reconstruction.
• Integrating spatial electrode information enhances BCI model effectiveness.