EEG-GAT: Graph Attention Networks for Classification of Electroencephalogram (EEG) Signals

Area of Science:

• Machine Learning
• Neuroscience
• Signal Processing

Background:

• Graph Neural Networks (GNNs) typically require a known graph topology.
• Real-world datasets, like electroencephalogram (EEG) data, often lack a predefined graph structure.
• Existing methods for GNNs in EEG analysis rely on hand-engineered graph structures.

Purpose of the Study:

• To develop a method for learning graph topology directly from EEG data.
• To extract subject-invariant features for cognitive task prediction.
• To overcome the limitation of predefined graph structures in GNN applications.

Main Methods:

• Parameterizing the graph shift operator (GSO) using a multi-head attention mechanism.
• Simultaneously learning unsupervised graph topology and graph convolutional kernel parameters.
• Applying the method to EEG data to explore functional neural connectivity during cognitive tasks.

Main Results:

• The multi-head attention mechanism effectively explores functional neural connectivity between EEG electrode sites.
• The unsupervised learning of graph topology is achieved in conjunction with GNN parameters.
• Demonstrated potential for subject-invariant feature extraction for cognitive task prediction.

Conclusions:

• The proposed method enables GNNs to learn graph topology from data, addressing a key limitation.
• This approach enhances the applicability of GNNs to complex, real-world datasets like EEG.
• It offers a promising direction for advancing brain-computer interfaces and cognitive neuroscience research.