MBGA-Net: A multi-branch graph adaptive network for individualized motor imagery EEG classification

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning has advanced Motor Imagery (MI) electroencephalography (EEG) signal classification accuracy.
• Current models struggle to achieve high classification accuracy for individual users.
• Precise recognition of individual EEG signals is critical for medical rehabilitation and intelligent control.

Purpose of the Study:

• To develop a novel deep learning model for precise individual MI EEG signal classification.
• To address the limitations of existing models in personalized EEG signal recognition.

Main Methods:

• Proposed a multi-branch graph adaptive network (MBGA-Net).
• Employed an adaptive technique to match individual EEG signals with suitable time-frequency processing methods based on spatio-temporal features.
• Utilized an enhanced attention mechanism and deep convolutional methods with residual connectivity within each model branch.

Main Results:

• MBGA-Net validated on BCI Competition IV datasets 2a and 2b.
• Dataset 2a: Average accuracy 87.49%, average kappa 0.83, with a low individual kappa standard deviation of 0.08.
• Dataset 2b: Average classification accuracies of 85.71%, 85.83%, and 86.99% across the three network branches.

Conclusions:

• MBGA-Net effectively classifies motor imagery EEG signals with strong generalization.
• The adaptive matching technique significantly enhances individual classification accuracy.
• The model shows promise for practical applications in EEG-based systems.