EEG-GMACN: Interpretable EEG Graph Mutual Attention Convolutional Network.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Electroencephalogram (EEG) records brain electrical activity, crucial for understanding neurological conditions and brain-computer interfaces.
• Graph Signal Processing (GSP) analyzes EEG by incorporating electrode topology, but lacks interpretability and confidence assessment.
• Current GSP methods for EEG analysis struggle with electrode importance and reliable prediction confidence.

Purpose of the Study:

• To develop a novel EEG analysis model that enhances interpretability and prediction confidence.
• To introduce an 'Inverse Graph Weight Module' for interpretable electrode importance in EEG classification.
• To improve the distinction of critical electrodes and assess prediction uncertainty in EEG analysis.

Main Methods:

• Proposed an EEG Graph Mutual Attention Convolutional Network (EEG-GMACN).
• Incorporated an 'Inverse Graph Weight Module' for interpretable electrode graph weights.
• Integrated a mutual attention mechanism and credibility calibration for enhanced analysis and uncertainty assessment.

Main Results:

• The EEG-GMACN model provides interpretable electrode importance, enhancing clinical credibility.
• The mutual attention mechanism effectively distinguishes critical electrodes in EEG data.
• Credibility calibration allows for reliable assessment of prediction uncertainty in EEG analysis.

Conclusions:

• The proposed EEG-GMACN enhances the transparency and effectiveness of EEG analysis.
• This approach improves the clinical credibility and interpretability of EEG classification results.
• The study paves the way for broader clinical and neuroscience research applications of EEG analysis.