Sleep staging from electrocardiography and respiration with deep learning

Area of Science:

• Cardiology
• Sleep Medicine
• Artificial Intelligence

Background:

• Sleep staging traditionally relies on electroencephalography (EEG), which is not always accessible.
• Heart rhythms and breathing patterns also reflect sleep states.

Purpose of the Study:

• To develop and validate deep neural networks for sleep staging using electrocardiogram (ECG) and respiratory signals.
• To assess the feasibility of sleep staging without EEG.

Main Methods:

• Utilized a dataset of 8682 polysomnograms.
• Trained five deep neural networks (CNNs and LSTMs) using ECG R-peak timing, abdominal/chest respiratory effort, and combinations thereof.
• Evaluated performance using Cohen's kappa for all five sleep stages and for awake/REM/NREM discrimination.

Main Results:

• ECG combined with abdominal respiratory effort yielded the best performance for all five sleep stages (Cohen's kappa = 0.585).
• Discriminating awake, REM, and NREM sleep achieved a Cohen's kappa of 0.760.
• Performance was better in younger individuals and robust across varying body mass index, apnea severity, and medications.

Conclusions:

• ECG and respiratory effort signals contain significant information for sleep staging in a diverse population.
• This non-EEG approach shows promise for sleep research and clinical applications where EEG is impractical.
• Deep learning models can effectively stage sleep using readily available physiological signals.