A Hierarchical Attention-Based Method for Sleep Staging Using Movement and Cardiopulmonary Signals

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Science

Background:

• Polysomnography (PSG) is the standard for sleep staging but is uncomfortable and costly.
• Automatic sleep staging using body movement and cardiopulmonary signals is an emerging alternative.
• Existing deep learning models like LSTM and CNN have limitations in sequence modeling for sleep staging.

Purpose of the Study:

• To develop a hierarchical attention-based deep learning method for improved automatic sleep staging.
• To utilize body movement, electrocardiogram (ECG), and abdominal breathing signals for sleep stage classification.
• To overcome the limitations of traditional LSTM and CNN models in sequence data processing.

Main Methods:

• A novel hierarchical attention-based deep learning architecture was developed.
• Multi-head self-attention was employed to model global context in feature sequences.
• The attention mechanism was coupled with Convolutional Neural Networks (CNN) for hierarchical weight assignment.
• The method was evaluated on two public sleep datasets.

Main Results:

• The proposed method achieved high performance in classifying three sleep stages.
• Accuracy reached 84.3%, F1 score was 0.8038, and Cohen's Kappa coefficient was 0.7036.
• The hierarchical self-attention mechanism demonstrated effectiveness in processing feature sequences for sleep staging.

Conclusions:

• The developed hierarchical attention-based deep learning method offers a promising non-invasive approach for sleep monitoring.
• This technique outperforms existing methods, providing accurate sleep stage classification.
• The findings pave the way for long-term sleep monitoring using readily available non-invasive sensors.