Filter bank sinc-convolutional network with channel self-attention for high performance motor imagery decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor Imagery Brain-Computer Interface (MI-BCI) is a key non-invasive BCI paradigm for identifying motor intentions.
• Deep learning, particularly lightweight networks, shows promise in MI-BCI but requires performance enhancement.

Purpose of the Study:

• To develop a high-performance MI-BCI decoding model.
• To improve the spatio-temporal feature extraction and selection for electroencephalography (EEG) data.
• To enhance the generalization capability of MI-BCI models.

Main Methods:

• Designed a filter bank with sinc-convolutional layers for spatio-temporal feature extraction.
• Introduced Channel Self-Attention for feature selection using global and local information.
• Developed a data augmentation method using multivariate empirical mode decomposition.

Main Results:

• Achieved high accuracies on three open MI datasets: 78.20% (4-class) on BCI Competition IV IIa, 87.34% (2-class) on BCI Competition IV IIb, and 72.03% (2-class) on OpenBMI.
• Significantly outperformed existing deep learning methods by at least 2.27% (p < 0.05).

Conclusions:

• The proposed Filter Bank Sinc-convolutional Network with Channel Self-Attention offers a novel and effective approach for MI-BCI decoding.
• This method provides a promising option for developing BCI systems for motor rehabilitation.