Temporally Continuous Automated Sleep-Wake Classification Using Deep Learning.

Area of Science:

• Sleep science
• Computational neuroscience
• Artificial intelligence in medicine

Background:

• Current sleep scoring relies on fixed 30-second epochs, which may not accurately represent sleep dynamics.
• This limitation can hinder precise physiological sleep assessment.

Purpose of the Study:

• To develop a deep learning-based sleep-wake classifier with high temporal resolution.
• To utilize temporally continuous manual scoring, bypassing fixed epoch boundaries.
• To improve the physiological consistency of sleep assessment.

Main Methods:

• A U-Net based deep learning model was trained on sleep-wake data.
• Transfer learning was employed, fine-tuning the model with temporally continuous scored data.
• The model was validated on independent datasets using continuous scoring.

Main Results:

• The classifier achieved high concordance (88.96% and 88.23%) with continuous manual scoring.
• Strong correlations were observed between 1-second predictions and manual scoring for total sleep time (r=0.93) and sleep-to-wake transitions (r=0.67).

Conclusions:

• The developed model effectively addresses limitations of traditional 30-second epoch scoring.
• This approach offers a practical foundation for more physiologically consistent sleep-wake assessment.
• Prediction confidence estimates can guide targeted review of potential misclassifications.