U-Sleep: resilient high-frequency sleep staging.

Area of Science:

• Sleep medicine
• Artificial intelligence in healthcare
• Computational neuroscience

Background:

• Sleep disorders are prevalent and linked to increased mortality.
• Accurate sleep staging from polysomnography (PSG) is crucial for clinical decisions.
• Manual sleep staging is labor-intensive and requires expert evaluation of EEG and EOG data.

Purpose of the Study:

• To introduce U-Sleep, a deep learning system for automated sleep staging.
• To evaluate U-Sleep's performance across diverse patient cohorts and PSG protocols.
• To demonstrate U-Sleep's potential for improved sleep analysis and clinical support.

Main Methods:

• Development of a fully convolutional neural network (U-Sleep) for automated sleep staging.
• Training and evaluation on PSG recordings from 15,660 participants across 16 clinical studies.
• Testing U-Sleep with various EEG/EOG channel combinations and at intervals shorter than 30 seconds.

Main Results:

• U-Sleep achieved performance comparable to state-of-the-art systems on multiple datasets.
• The system demonstrated accuracy on par with human experts when compared to consensus scores.
• U-Sleep successfully segmented sleep stages using arbitrary EEG/EOG channel combinations and shorter intervals.

Conclusions:

• U-Sleep provides accurate, automated sleep staging, supporting clinical workflows and reducing healthcare costs.
• The system offers potential for novel sleep analysis through high-resolution staging.
• U-Sleep can enhance diagnostic capabilities, especially where human expertise is limited.