MCAF-Net: Multi-Channel Temporal Cross-Attention Network with Dynamic Gating for Sleep Stage Classification

Area of Science:

• Biomedical Engineering
• Neuroscience
• Sleep Medicine

Background:

• Automated sleep stage classification is crucial for objective sleep evaluation and clinical diagnosis.
• Current methods primarily rely on single-channel electroencephalogram (EEG) signals, overlooking valuable information from other biosignals.
• Existing multi-channel fusion techniques often use simple concatenation, failing to capture complex cross-channel correlations.

Purpose of the Study:

• To develop a novel network architecture for enhanced sleep stage classification by effectively integrating multi-channel polysomnography (PSG) data.
• To address the limitations of conventional multi-channel fusion methods in modeling inter-channel dependencies.

Main Methods:

• Proposed MCAF-Net architecture utilizing temporal convolution modules for channel-specific feature extraction.
• Introduced a dynamic gated multi-head cross-channel attention mechanism (MCAF) to model interdependencies between physiological channels.
• Utilized multi-channel polysomnography (PSG) data for training and evaluation.

Main Results:

• MCAF-Net successfully integrated information from multiple physiological channels.
• The proposed method demonstrated significant improvements in sleep stage classification accuracy.
• Achieved superior performance compared to the majority of existing sleep stage classification algorithms.

Conclusions:

• The novel MCAF-Net architecture effectively models cross-channel correlations for improved sleep stage classification.
• Multi-channel data integration using advanced attention mechanisms offers a promising direction for objective sleep evaluation.
• MCAF-Net represents a significant advancement in automated sleep analysis using polysomnography data.