Adaptive graph convolutional neural network incorporating ECG for individualized motor imagery EEG classification

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery electroencephalogram (EEG) signal recognition is vital for medical rehabilitation and intelligent control.
• Existing cross-subject recognition methods struggle with inter-subject variability, limiting generalization.
• Reliable EEG signal prediction is crucial due to safety risks associated with misclassification.

Purpose of the Study:

• To develop adaptive modeling strategies for robust cross-subject recognition of EEG signals.
• To improve the generalization capabilities of EEG-based brain-computer interfaces.
• To address the challenge of substantial inter-subject variability in EEG data.

Main Methods:

• A Hybrid Adaptive Domain Graph Convolutional Network (HAD-GCN) was proposed.
• Multi-level adaptability was achieved through an adaptive generator synthesizing ECG signals from EEG and an adaptive splitter for time-frequency domain processing.
• The HAD-GCN model integrates spatial and temporal adaptive strategies for enhanced feature extraction.

Main Results:

• Cross-subject experiments on the Mixed dataset yielded 83.10% ± 6.54% accuracy and a Kappa of 0.778 ± 0.06.
• Experiments on the BCI Competition IV-2a dataset achieved 74.81% ± 8.97% accuracy and a Kappa of 0.655 ± 0.09.
• The HAD-GCN demonstrated significant improvements in cross-subject classification performance.

Conclusions:

• The HAD-GCN model significantly enhances cross-subject classification performance and prediction reliability.
• The multi-level adaptive approach demonstrates strong generalization capabilities for EEG-based technologies.
• The proposed method offers potential for practical applications in brain-computer interfaces and medical rehabilitation.