Optimizing automated sleep stage scoring of 5-s mini-epochs: a transfer learning study

Area of Science:

• Neuroscience
• Computational Biology
• Sleep Medicine

Background:

• Conventional sleep staging uses 30-second epochs, which may miss brief sleep stage changes.
• Previous research explored mini-epochs for more detailed sleep analysis.

Purpose of the Study:

• To optimize the U-Sleep deep learning model for 5-second mini-epoch sleep scoring.
• To evaluate if optimized U-Sleep achieves performance comparable to human scoring in mini-epochs.
• To enable more precise and detailed sleep characterization.

Main Methods:

• Created a dataset of 48,000 human-scored 5-second mini-epochs from 100 polysomnograms (PSGs).
• Applied transfer learning to optimize the U-Sleep model using human-scored mini-epochs.
• Assessed model performance using F1-scores, confusion matrices, stage distributions, and transition rates.

Main Results:

• Human-scored mini-epochs revealed significantly more transitions and different stage distributions compared to standard epochs.
• Optimized U-Sleep performance improved from an F1-score of 0.74 to 0.81.
• Optimized U-Sleep's stage distributions closely matched human-scored mini-epochs, with increased transition rates.

Conclusions:

• Optimized U-Sleep achieves high performance in 5-second mini-epoch scoring, comparable to traditional methods.
• Demonstrates the feasibility of precise, automated high-resolution sleep staging.
• Suggests future validation and application to full-night sleep recordings.