Expert-level sleep staging using an electrocardiography-only feed-forward neural network.

Area of Science:

• Sleep Medicine
• Neuroscience
• Biomedical Engineering

Background:

• Polysomnography (PSG) is the gold standard for sleep stage classification but is costly, invasive, and inconvenient.
• Current wearable sleep trackers lack the accuracy of PSG.
• Accurate sleep stage classification is vital for diagnosing sleep disorders and advancing neuroscience research.

Purpose of the Study:

• To develop an accurate, low-cost, and convenient method for sleep stage classification using only electrocardiography (ECG) data.
• To achieve performance comparable to PSG using a single ECG lead.
• To democratize access to high-quality sleep studies.

Main Methods:

• A feed-forward neural network was trained using a single lead of ECG data.
• A novel loss function was developed to optimize for Cohen's kappa.
• The model was validated on a large, diverse dataset spanning ages 5 to 90.

Main Results:

• The model achieved a median five-stage Cohen's kappa of 0.725, demonstrating non-inferior performance compared to human inter-rater agreement.
• The method provides an inexpensive, automated, and convenient alternative to PSG.
• Real-time scoring capability was developed.

Conclusions:

• Cardiosomnography, using only ECG, offers a viable, gold-standard-level alternative for sleep stage classification.
• This approach can significantly enhance sleep research and personalized healthcare by making high-quality sleep studies more accessible.
• The method has the potential to move expert-level sleep analysis from clinical settings into real-world environments.