Spatiotemporal convolution sleep network based on graph attention mechanism with automatic feature extraction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Signal Processing

Background:

• Graph neural networks (GNNs) are prevalent in automatic sleep staging.
• Existing GNNs often rely on computationally expensive spectral methods.
• There is a need for more efficient and flexible GNN approaches for sleep staging.

Purpose of the Study:

• To introduce a non-spectral graph attention network approach for sleep staging.
• To develop a spatiotemporal convolution sleep network (ST-GATv2) for enhanced feature extraction and generalization.
• To improve the accuracy and efficiency of automatic sleep staging.

Main Methods:

• Utilized graph attention networks v2 (GATv2) for spatial information extraction (S-GATv2).
• Implemented multi-convolutional layers for automatic feature extraction.
• Applied graph attention to the time domain (T-GATv2) and introduced a modified function to capture temporal dynamics and trends.

Main Results:

• The ST-GATv2 model achieved a highest accuracy of 89.0% on the SS3 dataset.
• The T-GATv2 block and modified function contributed to an approximate 0.5% improvement in Kappa and F1-score.
• The proposed model demonstrated superior performance compared to other advanced models.

Conclusions:

• Graph attention mechanisms and novel blocks (T-GATv2, modified function) show significant potential in sleep classification.
• The ST-GATv2 model is proposed as an effective tool for sleep staging in both healthy and diseased individuals.
• The non-spectral, spatiotemporal approach offers a more flexible and intuitive alternative to spectral GNNs.