High Classification Accuracy Does Not Guarantee Neurophysiological Validity: A Critical Evaluation of Graph Attention Network Interpretability in MEG

Decoding task-specific magnetoencephalography (MEG) signals is promising for advancing brain-computer interfaces (BCIs)and understanding human cognition. While deep learning models have shown competitive classification performance on MEG datasets, the neurophysiological validity of their learned representations remains underexplored. In this study, we propose a novel approach combining a Graph Attention Network (GAT) architecture optimized with Particle Swarm Optimization (PSO)to classify MEG signals from a multimodal face perception dataset. Beyond evaluating classification accuracy, we critically assess whether the model’s learned sensor importance aligns with source-localized neural activation patterns derived via dynamic statistical parametric mapping (dSPM) and LCMV beamforming techniques. Quantitative comparisons based on cosine similarity and Pearson correlation analyses reveal a critical dissociation between superficial directional alignment and meaningful neurophysiological correspondence. While cosine similarity values remain consistently high across subjects(0.66-0.81), indicating broad spatial agreement, Pearson correlations are weak or negative (-0.28 to +0.24), demonstrating poor fine-grained pattern matching. Furthermore, analysis of peak coordinate locations shows substantial spatial misalignment(15.2-89.7 mm Euclidean distance) between GAT saliency maxima and LCMV activation peaks. Our results provide critical evidence that high classification accuracy does not guarantee neurophysiological interpretability in graph neural networks applied to MEG data.