SFE-GAT: Structure-Feature Evolution Graph Attention Network for Motor Imagery Decoding

Area of Science:

• Neuroscience
• Machine Learning
• Brain-Computer Interfaces

Background:

• Current motor imagery EEG decoding uses static functional connectivity, failing to capture dynamic brain network changes.
• This limitation hinders accurate decoding and understanding of neural processes during tasks.

Purpose of the Study:

• Develop a graph neural network to simulate neurodynamic processes for improved EEG decoding.
• Provide computational insights into the stage-wise reorganization of brain networks.

Main Methods:

• Propose the Structure-Feature Evolution Graph Attention Network (SFE-GAT) with an inter-layer evolution mechanism.
• Dynamically co-adapt graph topology and node features using a graph autoencoder with Monte Carlo sampling.
• Initialize with phase-locking value connectivity and spectral features.

Main Results:

• Achieved 77.70% (subject-dependent) and 66.59% (subject-independent) accuracy on the BCI Competition IV-2a dataset.
• Outperformed existing baseline models in EEG decoding.
• Observed sparsification and strengthening of task-critical connections in evolved graphs, suggesting hierarchical processing.

Conclusions:

• SFE-GAT advances EEG decoding by employing a dynamic graph architecture.
• The model offers a computational framework for studying hierarchical organization in motor cortex activity.
• Links adaptive graph learning with neural dynamics for enhanced brain-computer interfaces.