DistillSleep: real-time, on-device, interpretable sleep staging from single-channel electroencephalogram

Area of Science:

• Neuroscience and Biomedical Engineering
• Artificial Intelligence in Healthcare
• Sleep Medicine Technology

Background:

• Polysomnography (PSG) is the gold standard for sleep staging but is resource-intensive.
• Current methods require laboratory settings, multiple sensors, and manual scoring.
• There is a need for accessible, real-time sleep staging solutions.

Purpose of the Study:

• To develop DistillSleep, a single-channel electroencephalogram (EEG) framework for accurate, real-time, and interpretable sleep staging.
• To enable sleep staging on resource-constrained devices.
• To create a portable and efficient alternative to traditional PSG.

Main Methods:

• Developed a DistillSleep framework with a high-capacity teacher model and a compact student model for edge deployment.
• Utilized a Multi-Wavelength Pyramid module and Transformer-based architecture for feature extraction.
• Employed knowledge distillation to transfer expertise from the teacher to the student model.
• Trained and evaluated on over 10,000 overnight recordings from six diverse cohorts.

Main Results:

• The teacher model achieved state-of-the-art Macro-F1 scores across multiple cohorts.
• The student model maintained competitive accuracy (up to 79.7% Macro-F1) with millisecond-scale latency on embedded platforms.
• DistillSleep demonstrated a 115-fold reduction in computational load compared to prior methods.
• Interpretability features, including frequency-band importance and inter-epoch context, were successfully transferred to the student model.

Conclusions:

• DistillSleep provides expert-level accuracy, low latency, and transparent decision-making in a single-channel EEG format.
• The framework is suitable for point-of-care diagnostics, therapy titration, and large-scale home monitoring.
• DistillSleep expands the reach of sleep medicine by offering a practical and trustworthy solution.