SleepPPG-Net: A Deep Learning Algorithm for Robust Sleep Staging From Continuous Photoplethysmography.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Medicine

Background:

• Sleep staging is crucial for diagnosing sleep disorders and managing sleep health.
• Traditional sleep staging is labor-intensive, requiring manual labeling in clinical settings.
• Automated sleep staging using physiological signals is an active area of research.

Purpose of the Study:

• To develop and validate a deep learning model for robust 4-class sleep staging using raw photoplethysmography (PPG) time series.
• To evaluate the performance of the developed model against state-of-the-art algorithms.
• To explore the potential of PPG-based sleep staging for clinical applications and wearable devices.

Main Methods:

• Utilized two public sleep databases with 2,374 patients and 23,055 hours of raw PPG recordings.
• Developed SleepPPG-Net, a deep learning model employing residual convolutional and temporal convolutional networks.
• Trained the model end-to-end for automatic feature extraction and capturing long-range contextual information.

Main Results:

• SleepPPG-Net achieved a median Cohen's Kappa (κ) score of 0.75, outperforming the best state-of-the-art (SOTA) approach (κ=0.69).
• Demonstrated good generalization to an external database, achieving a κ score of 0.74 after transfer learning.
• The model achieved new SOTA performance in 4-class sleep staging from PPG signals.

Conclusions:

• SleepPPG-Net offers a novel and highly accurate method for automated sleep staging using PPG.
• The model's performance meets requirements for clinical applications, including obstructive sleep apnea diagnosis and monitoring.
• This research facilitates the development of advanced wearable devices for sleep health management.