Classification of consciousness disorders based on graph convolution and attention mechanism

This study develops an interpretable, task-adaptive EEG diagnostic framework to overcome the subjectivity and time burden of behavioral scales like the CRS-R. The model is designed to learn task-relevant inter-electrode connectivity directly from data via a trainable adjacency matrix within a graph-convolutional architecture, making it robust to individual and state-related variability. It also integrates a channel-time dual-dimensional Efficient Channel Attention mechanism to jointly model spatial and temporal dependencies. Interpretability is emphasized by extracting model-derived channel saliency to identify a compact set of clinically informative electrodes and by testing whether these reduced-channel subsets can sustain high diagnostic performance. Clinical applicability is prioritized by validating stability with cross-validation and by targeting bedside deployment through channel reduction and streamlined decoders, while laying groundwork for physiological interpretation and multi-center, multimodal validation. We propose GCENet, which uses three stacked GCN layers with a learnable adjacency matrix and ECA modules to dynamically weight channel and temporal features. Data comprised 119 EEG segments recorded with a 20-channel system; performance was assessed using ten-fold cross-validation. GCN-ECA achieved mean accuracy of 87.12% and AUC of 92.76%, outperforming baselines. The learned connectivity emphasized frontal and occipital channels, supporting an interpretable channel-reduction strategy. This attention-enhanced GCN offers an objective, scalable alternative to behavioral assessment and a practical path toward reduced-channel, bedside DOC monitoring and broader clinical validation.