A Novel Recognition and Classification Approach for Motor Imagery Based on Spatio-Temporal Features

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Electroencephalography (EEG) signals in motor imagery are non-stationary with low signal-to-noise ratios, posing challenges for brain-computer interfaces (BCIs).
• Accurate feature extraction from motor imagery EEG is crucial for effective BCIs in medical rehabilitation.

Purpose of the Study:

• To propose a novel motor imagery EEG signal classification model.
• To enhance the accuracy and robustness of BCIs for medical rehabilitation applications.

Main Methods:

• Constructing functional brain networks using various brain functional connectivity metrics.
• Calculating graph theory features to analyze brain network characteristics during different motor tasks.
• Employing graph convolutional networks (GCNs) for the classification of motor imagery tasks based on functional brain networks.

Main Results:

• The proposed model achieved 88.39% accuracy in multi-subject classification on the Physionet dataset, surpassing traditional methods.
• In single-subject conditions, the model reached an average accuracy of 99.31%, effectively handling individual variability.
• Analysis revealed significantly higher functional connectivity strength during the 'both fists' task compared to other motor imagery tasks.

Conclusions:

• The combined functional brain network and GCN model demonstrates superior performance in motor imagery classification.
• The findings offer new insights into functional connectivity patterns across different motor tasks and brain regions.
• This approach holds significant potential for advancing BCIs in medical rehabilitation.