A Residual Based Attention Model for EEG Based Sleep Staging.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Manual sleep staging is crucial for diagnosing sleep disorders but is time-consuming and subjective.
• Automated sleep staging methods, particularly deep learning, show promise but often neglect EEG domain knowledge and struggle with temporal context.
• Existing recurrent neural networks like LSTM are inefficient for modeling global temporal information in EEG.

Purpose of the Study:

• To develop a more accurate and efficient automated sleep staging system.
• To incorporate EEG frequency band domain knowledge into deep learning models.
• To improve the modeling of global temporal context in sleep staging.

Main Methods:

• A multi-scale deep architecture was proposed, decomposing EEG signals into frequency bands for CNN input.
• Transformer's multi-head self-attention module was utilized to capture global temporal context.
• A residual-based architecture was employed for end-to-end training.

Main Results:

• The proposed method demonstrated effectiveness on the MASS and sleep-EDF datasets.
• Significant efficiency gains were observed, with up to 12 times less training time compared to state-of-the-art methods.
• The multi-scale approach combined with self-attention improved sleep staging performance.

Conclusions:

• The novel multi-scale deep architecture effectively utilizes EEG frequency domain knowledge and global temporal context for improved sleep staging.
• The method offers a significant reduction in training time, making automated sleep staging more practical.
• This approach advances automated sleep staging, aiding in the diagnosis and treatment of sleep disorders.