AnySleep: a channel-agnostic deep learning system for high-resolution sleep staging in multi-center cohorts

Area of Science:

• Neuroscience
• Biomedical Engineering
• Sleep Medicine

Background:

• Manual sleep staging from polysomnography (PSG) is labor-intensive and poses challenges for multi-center studies.
• Traditional 30-second epoch scoring lacks physiological basis and limits biomarker discovery on shorter timescales.
• Variability in electrode count and montage across centers complicates harmonized sleep research.

Purpose of the Study:

• To develop a deep neural network model (AnySleep) for automated sleep scoring using diverse electroencephalography (EEG) and electrooculography (EOG) data.
• To enable sleep scoring at adjustable temporal resolutions, including sub-30-second intervals.
• To facilitate robust generalization across multiple clinical sites and heterogeneous data acquisition setups.

Main Methods:

• Trained and validated AnySleep on over 19,000 overnight recordings from 21 diverse datasets, totaling nearly 200,000 hours of EEG and EOG data.
• Evaluated model performance against established baselines at 30-second epochs and on sub-30-second timescales.
• Assessed model's ability to perform with varying numbers of channels, including limited or absent EOG and single EEG derivations.

Main Results:

• AnySleep achieved state-of-the-art performance, matching or exceeding traditional methods at 30-second epochs.
• Model performance remained strong even with minimal or absent EOG or single EEG channels.
• On sub-30-second timescales, AnySleep accurately identified short wake intrusions (arousals) and improved prediction of physiological and pathophysiological conditions.

Conclusions:

• AnySleep offers a flexible and robust solution for automated sleep scoring using readily available EEG/EOG data.
• The model's ability to handle data heterogeneity and variable temporal resolutions facilitates large-scale, multi-center sleep research.
• Public availability of AnySleep is expected to accelerate the discovery of novel sleep biomarkers and improve clinical sleep analysis.