InsightSleepNet: the interpretable and uncertainty-aware deep learning network for sleep staging using continuous...

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Signal Processing

Background:

• Addressing limitations of current sleep monitoring methods, including inconvenience and high costs.
• Utilizing continuous photoplethysmography (PPG) signals from wearable devices for sleep monitoring.
• Aiming for an efficient sleep staging method with interpretable and uncertainty-aware predictions to aid medical professionals.

Purpose of the Study:

• To develop and validate InsightSleepNet, a novel 4-class sleep staging model.
• To enhance the interpretability and uncertainty estimation of sleep staging predictions.
• To provide a reliable tool for medical professionals supporting clinical decision-making.

Main Methods:

• Developed a 4-class sleep staging model incorporating local attention, InceptionTime, time-distributed dense layer, TCN, and CNN.
• Employed a local attention module for epoch-wise PPG data impact analysis via TCN.
• Utilized energy score estimation for uncertainty quantification and selective prediction.

Main Results:

• InsightSleepNet demonstrated improved performance across MESA, CFS, and CAP datasets after applying an energy score threshold.
• Accuracy on MESA improved to 84.8-86.1%, Cohen's kappa to 0.75-0.78, and weighted F1 to 0.848-0.861.
• Significant performance gains were also observed on CFS and CAP datasets, with identified correlations between PPG signal peaks and sleep stage classification.

Conclusions:

• InsightSleepNet offers a 4-class sleep staging solution using continuous PPG data for wearable devices.
• The model enhances sleep analysis interpretability, aiding medical professionals in intervention-based predictions and clinical decisions.
• Provides a reliable second opinion for medical settings, improving the decision-making process.