Classification of consciousness disorders based on graph convolution and attention mechanism

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Behavioral scales for diagnosing disorders of consciousness (DOC) are subjective and time-consuming.
• Existing electroencephalography (EEG) analysis methods often lack interpretability and adaptability.
• Objective and scalable diagnostic tools are needed for clinical settings.

Purpose of the Study:

• To develop an interpretable, task-adaptive EEG diagnostic framework for disorders of consciousness (DOC).
• To overcome the subjectivity and time burden associated with traditional behavioral assessment scales.
• To create a practical, reduced-channel system for bedside monitoring.

Main Methods:

• Developed GCENet, a graph convolutional network (GCN) with a learnable adjacency matrix and Efficient Channel Attention (ECA) modules.
• Utilized 119 EEG segments from a 20-channel system, assessed via ten-fold cross-validation.
• Employed channel saliency extraction for interpretability and channel reduction.

Main Results:

• GCENet achieved 87.12% mean accuracy and 92.76% AUC, outperforming baseline models.
• Learned inter-electrode connectivity highlighted frontal and occipital regions, supporting channel reduction.
• Reduced-channel subsets maintained high diagnostic performance, demonstrating clinical applicability.

Conclusions:

• The attention-enhanced GCN (GCENet) provides an objective and scalable alternative to behavioral assessments for DOC.
• The framework enables interpretable EEG analysis and supports reduced-channel, bedside monitoring.
• GCENet lays the groundwork for broader clinical validation and physiological interpretation.