Obstructive sleep apnea event detection using explainable deep learning models for a portable monitor

Area of Science:

• Artificial Intelligence in Healthcare
• Sleep Medicine Technology
• Biomedical Signal Processing

Background:

• Polysomnography (PSG) is the standard for obstructive sleep apnea (OSA) diagnosis but has significant limitations for widespread use.
• Portable monitors (PMs) integrated with deep learning (DL) offer a streamlined approach to OSA detection.

Purpose of the Study:

• To develop and evaluate deep learning models for detecting OSA events and calculating the apnea-hypopnea index (AHI) on portable monitors.
• To enhance the interpretability of deep learning models using techniques like Gradient-weighted Class Activation Mapping (Grad-CAM).

Main Methods:

• Trained multiple deep learning models on polysomnography data from the National Sleep Research Resource (NSRR).
• Optimized deep learning architecture hyperparameters for improved performance.
• Applied Grad-CAM to analyze model decision-making processes and identify relevant signal regions.

Main Results:

• A one-dimensional convolutional neural network (1D-CNN) achieved the best performance with 84.3% accuracy.
• Model interpretability analysis highlighted critical signal regions influencing OSA detection.
• The developed DL model demonstrated effectiveness in detecting OSA events and calculating AHI.

Conclusions:

• Explainable deep learning models show significant potential for portable monitors as an alternative to PSG for OSA detection.
• Further development is needed, but this approach represents a promising future direction for at-home sleep apnea diagnostics.