A Two-Stream Graph Convolutional Network Based on Brain Connectivity for Anesthetized States Analysis.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Clinical consciousness monitoring relies heavily on anesthesia state detection.
• Traditional methods lack analysis of topological changes in brain activity during anesthesia.
• Electrocorticography (ECoG) offers rich data for investigating neural mechanisms of anesthesia.

Purpose of the Study:

• To develop an efficient anesthetized state detection method using ECoG signals.
• To address the limitations of traditional monitoring by incorporating topological brain network changes.
• To propose a novel framework for enhanced clinical consciousness monitoring.

Main Methods:

• A two-stream graph convolutional network (GCN) framework was developed, utilizing one stream for topological structure features and another for node features.
• GCN Model 1 employed phase lag index (PLI) to construct brain connectivity networks and a dual-graph method for structure features.
• GCN Model 2 utilized average absolute signal amplitudes as node features, with a fully connected matrix for adjacency.

Main Results:

• The proposed GCN framework achieved high accuracy in distinguishing between awake, moderate, and deep sedation states.
• Group-level experiments reported an accuracy of 92.75%, with subject-level experiments achieving a mean accuracy of 93.50%.
• The method demonstrated effectiveness in learning both topological structure and node features from ECoG data.

Conclusions:

• The study validates the efficacy of graph convolutional networks for anesthesia monitoring.
• High recognition accuracy suggests that brain networks contain neurological markers indicative of anesthesia.
• The developed framework offers a promising advancement for clinical consciousness monitoring.