MS-Net: Sleep apnea detection in PPG using multi-scale block and shadow module one-dimensional convolutional neural...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Medicine

Background:

• Sleep Apnea (SA) is a common sleep disorder.
• Traditional polysomnography for SA detection is complex and not user-friendly for home monitoring.
• Photoplethysmography (PPG) offers a low-cost, convenient alternative for widespread SA screening.

Purpose of the Study:

• To develop and validate a novel deep learning model for accurate sleep apnea detection using PPG signals.
• To enhance the portability and accuracy of SA detection models for home-use applications.
• To address class imbalance issues in SA detection datasets.

Main Methods:

• A dual-channel input model combining a multi-scale one-dimensional convolutional neural network (MCNN) and a shadow one-dimensional convolutional neural network (SCNN).
• Extraction of time-series features using multi-scale temporal structures and utilization of redundant information via a shadow module.
• Implementation of balanced bootstrapping and class weighting to mitigate class imbalance.

Main Results:

• Achieved 82.0% average accuracy, 74.4% average sensitivity, and 85.1% average specificity for per-segment SA detection.
• Reached 93.6% average accuracy for per-recording SA detection following 5-fold cross-validation.
• Demonstrated good robustness and improved model portability.

Conclusions:

• The proposed MCNN-SCNN model provides an effective and robust method for sleep apnea detection using PPG signals.
• This approach is suitable for convenient and accurate SA screening in household settings.
• The developed model shows promise as a valuable tool for at-home sleep apnea diagnosis support.