Convolution-and Attention-Based Neural Network for Automated Sleep Stage Classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Polysomnography (PSG) is crucial for sleep health evaluation.
• Manual sleep stage scoring from PSG data is time-consuming and requires expert interpretation.
• Existing methods often lack interpretability and struggle with complex signal characteristics.

Purpose of the Study:

• To develop an interpretable deep learning model for automated sleep staging.
• To improve the efficiency and accuracy of PSG data analysis.
• To incorporate attention mechanisms for better feature extraction in sleep staging.

Main Methods:

• A novel neural network architecture combining Convolutional Neural Networks (CNNs) and attention mechanisms was proposed.
• The CNN component learns local signal characteristics within sleep epochs.
• The attention mechanism captures inter- and intra-epoch dependencies for enhanced feature learning.

Main Results:

• The model achieved high accuracies of 93.7% and 82.8% on the public sleep-edf and sleep-edfx databases.
• Macro-average F1-scores reached 84.5% and 77.8% respectively, demonstrating robust performance.
• The proposed model outperformed recently reported machine learning-based methods in sleep staging tasks.

Conclusions:

• The developed deep learning model offers an effective and interpretable solution for automatic sleep staging.
• The integration of CNNs and attention mechanisms significantly enhances the accuracy of sleep analysis.
• This approach promises to streamline sleep health evaluations by automating a critical, labor-intensive process.