A temporal multi-scale hybrid attention network for sleep stage classification.

Area of Science:

• Sleep Medicine
• Artificial Intelligence
• Biomedical Engineering

Background:

• Sleep is vital for overall health, and accurate sleep stage classification is essential for diagnosing sleep disorders.
• Current automatic sleep staging methods often fail to capture the nuances of sleep stage transitions or align with expert visual assessments.
• Polysomnography (PSG) is the gold standard for sleep monitoring, providing rich data for analysis.

Purpose of the Study:

• To develop an advanced deep learning model for automatic sleep stage classification from PSG data.
• To address the limitations of existing methods in handling sleep stage transitions and expert concordance.
• To enhance the accuracy and reliability of automated sleep analysis.

Main Methods:

• Proposing a novel Temporal Multi-scale Hybrid Attention Network (TMHAN) for sleep staging.
• Incorporating a temporal multi-scale mechanism to capture both short-term abrupt and long-term periodic transitions in PSG epochs.
• Utilizing a hybrid attention mechanism with 1-D local, 2-D global, and 2-D contextual sparse multi-head self-attention for sequence representations.
• Training an end-to-end model using a softmax layer on concatenated representations.

Main Results:

• TMHAN demonstrated superior performance compared to several baseline methods on two benchmark sleep datasets.
• The model effectively captured complex sleep stage transitions, aligning better with expert interpretations.
• Experimental results validated the effectiveness and accuracy of the proposed TMHAN model.

Conclusions:

• The developed TMHAN model offers a significant advancement in automatic sleep staging from PSG.
• The model's ability to integrate temporal dynamics and attention mechanisms improves classification performance.
• This work contributes to bridging the gap between deep learning techniques and clinical sleep medicine practices.