Exploring EEG Connectivity in Higher-Order Cognitive Tasks with Explainable Graph Neural Networks

This study investigates EEG connectivity during a higher-order cognitive task-spatial perspective-taking-using Graph Neural Networks (GNNs). We address the explainability limitations of EEGNet, a widely adopted model that struggles with interpretability tools to identify channel connectivity. In contrast, the proposed GNN architecture enhances interpretability while maintaining performance comparable to EEGNet. Our approach leverages a GNN based on Self-supervised Graph Attention Networks, incorporating a fully connected graph structure with channel embeddings to ensure a comprehensive representation of EEG connectivity, along with a convolutional encoder. Through visualization methods, we identify critical subgraphs corresponding to brain sub-networks that influence decision-making, emphasizing the pivotal role of the fronto-parietal network in spatial perspective-taking, especially under cognitively demanding conditions. Additionally, our findings demonstrate that conventional connectivity calculations based on correlation predominantly capture short-range interactions, overlooking critical long-range connections. In contrast, the proposed GNN approach effectively identifies key functional connections. These results highlight the potential of GNN-based methods in EEG analysis for complex cognitive tasks and underscore the importance of model interpretability in advancing neuroscientific research.