GCNs-Net: A Graph Convolutional Neural Network Approach for Decoding Time-Resolved EEG Motor Imagery Signals

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) signal analysis for brain-computer interfaces (BCI) traditionally ignores electrode topology.
• Neuroscience highlights the importance of network patterns in brain dynamics, suggesting Euclidean structures are insufficient.
• Existing methods lack the ability to fully capture complex inter-electrode signal interactions.

Purpose of the Study:

• To develop a novel deep learning (DL) framework to enhance the decoding performance of raw EEG signals for motor imagery (MI) tasks.
• To incorporate the functional topological relationships among EEG electrodes into the decoding process.
• To improve the accuracy and robustness of BCI systems.

Main Methods:

• A novel deep learning (DL) framework based on graph convolutional neural networks (GCNs) was developed.
• The graph Laplacian of EEG electrodes was constructed using the absolute Pearson's correlation matrix.
• A GCNs-Net architecture with graph convolutional, pooling, and fully-connected layers was employed for feature extraction and classification.

Main Results:

• The GCNs-Net achieved high average accuracies: 93.06% (subject) and 88.57% (group) on the PhysioNet dataset.
• On the high gamma dataset, accuracies reached 96.24% (subject) and 80.89% (group).
• The approach demonstrated adaptability, robustness to individual variability, and reproducible performance across cross-validation.

Conclusions:

• The GCNs-Net effectively decodes relevant features for brain MI by filtering EEG signals based on functional topological relationships.
• This novel framework represents a significant advancement toward more effective and efficient BCI systems.
• The DL library for EEG task classification is open-source, facilitating further scientific research.