An automated heart rate-based algorithm for sleep stage classification: Validation using conventional polysomnography...

Area of Science:

• Biomedical Engineering
• Sleep Science
• Artificial Intelligence

Background:

• Wearable devices enable sleep monitoring outside clinical settings, but often lack validation.
• Current systems are not always considered reliable alternatives to polysomnography (PSG).
• Independent validation is crucial for robustness and replicability of sleep staging algorithms.

Purpose of the Study:

• To validate a novel heart rate (HR)-based deep learning algorithm for sleep staging.
• To assess the algorithm's accuracy and precision across different sleep classification levels (2, 3, and 4 levels).
• To evaluate the algorithm's performance on both open-source and proprietary datasets.

Main Methods:

• A deep learning algorithm using heart rate (HR) data was developed for sleep staging.
• The algorithm was validated on the Physionet CinC dataset (n=994) and a proprietary Z3Pulse dataset (n=52).
• Performance metrics included Accuracy, Cohen's kappa, Sensitivity, Specificity, PPV, and NPV.

Main Results:

• The algorithm achieved high accuracy (up to 0.88) and Cohen's kappa (up to 0.61) in both datasets.
• The 4-level classification showed lower sensitivity but high specificity for Deep sleep.
• Performance was not affected by AHI or sex, but decreased with age.

Conclusions:

• HR-based sleep staging using deep learning is feasible and accurate.
• The algorithm, combined with inexpensive HR devices, offers a cost-effective, non-invasive home monitoring solution.
• The system demonstrates robustness across age, sex, and AHI scores, providing a viable alternative to PSG.