Multi-Layer Graph Attention Network for Sleep Stage Classification Based on EEG.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Neuroscience

Background:

• Graph neural networks (GNNs) show promise in sleep stage classification.
• Challenges remain in utilizing EEG channel interactions and extracting features from transitional sleep stages.

Purpose of the Study:

• To propose a Multi-layer Graph Attention Network (MGANet) to address limitations in current GNN-based sleep stage classification.
• To improve the accuracy and robustness of sleep stage classification models.

Main Methods:

• Developed MGANet incorporating node-level attention for channel interaction analysis in time-frequency and spatial domains.
• Employed a multi-head spatial-temporal mechanism for dynamic channel feature adjustment.
• Utilized stage-level attention to refine classification of easily confused sleep stages.

Main Results:

• MGANet achieved superior classification accuracy, MF1, and Kappa scores on ISRUC and SHHS datasets compared to state-of-the-art baselines.
• Achieved accuracy of 0.825 (ISRUC) and 0.814 (SHHS).
• Achieved MF1 scores of 0.775 (ISRUC) and 0.801 (SHHS).
• Achieved Kappa scores of 0.873 (ISRUC) and 0.827 (SHHS).

Conclusions:

• MGANet effectively utilizes epoch information from adjacent EEG channels.
• The proposed model enhances feature extraction for complex sleep transitions.
• MGANet demonstrates significant improvements in sleep stage classification accuracy.