FB-CGANet: filter bank channel group attention network for multi-class motor imagery classification

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Motor imagery-based brain-computer interfaces (MI-BCI) decode limb intentions from electroencephalography (EEG) signals.
• Lightweight neural networks offer efficient MI decoding but require performance enhancement for multi-class classification.

Purpose of the Study:

• To design a novel lightweight neural network, Filter Bank CGA Network (FB-CGANet), to improve multi-class motor imagery decoding accuracy.
• To investigate the efficacy of a hybrid filter bank and channel group attention (CGA) mechanism for enhanced feature extraction.

Main Methods:

• Developed FB-CGANet, integrating a hybrid filter bank for time-frequency information extraction and CGA for channel attention.
• Introduced a band exchange data augmentation technique tailored for filter bank structures to increase training data diversity.

Main Results:

• FB-CGANet achieved superior 4-class average accuracy (79.4%) on unseen data from the BCI Competition IV IIa dataset compared to existing methods.
• Cross-validation experiments demonstrated significantly higher average accuracy (93.5%) with the proposed FB-CGANet.

Conclusions:

• The study highlights the effectiveness of channel attention and filter bank structures within lightweight neural networks for MI-BCI.
• FB-CGANet presents a promising new method for improving multi-class motor imagery classification in brain-computer interfaces.