[A hybrid attention temporal sequential network for sleep stage classification].

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Sleep stage classification is crucial for diagnosing sleep disorders.
• Traditional methods (manual, machine learning) suffer from low efficiency and poor generalization.
• Deep neural networks show promise but often neglect intra-temporal and inter-channel information.

Purpose of the Study:

• To develop an advanced deep learning model for accurate automatic sleep stage classification.
• To address limitations of existing models by incorporating temporal sequential information and channel correlations.
• To improve the efficiency and generalization of sleep stage classification.

Main Methods:

• Proposed a hybrid attention temporal sequential network model.
• Utilized recurrent neural networks (RNNs) for temporal feature extraction from polysomnography.
• Implemented intra-temporal and channel attention mechanisms for feature fusion.
• Employed an inter-temporal attention mechanism for contextual representation fusion.

Main Results:

• The model achieved high accuracy (up to 0.801) and macro average F1-scores (up to 0.752) on benchmark datasets.
• Demonstrated superior performance compared to ten state-of-the-art baseline methods.
• Effectively integrated intra-temporal, channel-correlated, and inter-temporal representations.

Conclusions:

• The proposed hybrid attention temporal sequential network offers an effective end-to-end solution for automatic sleep stage classification.
• The model's ability to capture complex sequential and correlational data patterns enhances classification performance.
• This approach represents a significant advancement in sleep medicine diagnostics through improved automated analysis.