Optimising Sleep Stage Detection Using a Minimal Non-EEG Physiological Signal Set and Deep Learning.

Area of Science:

• Biomedical Engineering
• Sleep Science
• Artificial Intelligence

Background:

• Automatic sleep stage classification is crucial for at-home monitoring but often relies on electroencephalogram (EEG) signals, limiting reproducibility.
• Current methods using EEG are complex and hinder widespread, non-invasive application.

Purpose of the Study:

• To develop a reproducible engineering framework for a deep learning model for sleep stage classification.
• To enable accurate sleep monitoring using easily acquired physiological signals, avoiding EEG.

Main Methods:

• A U-Net based deep learning model was developed for classifying sleep into five or four stages.
• The model utilized oxygen saturation (SpO2), heart rate (HR), and abdominal respiratory effort (AbdRes), predicting on a per-second basis.
• Model optimization was performed using Keras Tuner with the Hyperband algorithm on the SHHS2 dataset and validated on the MESA dataset.

Main Results:

• The model achieved weighted F1-scores of 68% (five-stage) and 71% (four-stage) with Cohen's Kappa of 0.61 and 0.67 on the SHHS2 dataset.
• Consistent performance was observed during external validation on the MESA dataset.
• The study demonstrated strong generalization capabilities of the developed model.

Conclusions:

• The lightweight, EEG-free approach offers a practical and scalable solution for clinically relevant sleep monitoring.
• This framework advances non-invasive, at-home sleep analysis, overcoming limitations of traditional methods.
• The per-second prediction capability enhances the granularity of sleep stage classification.