Detection of sleep apnea from single-channel electroencephalogram (EEG) using an explainable convolutional neural...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Neuroscience

Background:

• Sleep apnea (SA) is a prevalent disorder causing breathing cessation during sleep, leading to significant health issues.
• Current diagnosis relies on polysomnography, often facing long waitlists, delaying critical treatment.
• Automated and simplified SA detection methods are needed to improve diagnostic accessibility.

Purpose of the Study:

• To develop and validate an explainable convolutional neural network (CNN) for detecting sleep apnea events using single-channel electroencephalography (EEG).
• To assess the generalizability of the CNN across different subjects.
• To elucidate the underlying mechanisms of the CNN's detection capabilities.

Main Methods:

• An explainable CNN with three convolutional layers was designed for SA detection from single-channel EEG.
• Hyperparameter tuning was performed using the Hyperband algorithm, and optimization used Adam.
• Network performance was evaluated using subject-wise 10-fold cross-validation and explained using critical-band masking (CBM) and filter kernel lesioning.

Main Results:

• The developed CNN achieved 69.9% accuracy and a Matthews correlation coefficient (MCC) of 0.38 in detecting SA events.
• Critical-band masking and lesioning analyses revealed the network learned frequency-band information correlating with known SA biomarkers (delta and beta bands).
• The model demonstrated generalizability across subjects.

Conclusions:

• Single-channel EEG holds clinical potential for the simplified and automated detection of sleep apnea.
• Explainable AI methods can validate the biological relevance of AI-driven diagnostic tools.
• This approach may help reduce diagnostic wait times and improve patient outcomes for sleep apnea.