SleepECG-Net: Explainable Deep Learning Approach With ECG for Pediatric Sleep Apnea Diagnosis

Area of Science:

• Cardiology
• Pediatrics
• Artificial Intelligence

Background:

• Obstructive sleep apnea (OSA) in children is common and linked to heart problems.
• Current diagnosis via polysomnography is complex and inaccessible, leading to underdiagnosis.
• Existing deep learning (DL) models for OSA lack interpretability.

Purpose of the Study:

• Introduce an interpretable DL model, SleepECG-Net, for estimating pediatric OSA severity.
• Utilize explainable AI (XAI) to identify ECG patterns associated with OSA and cardiovascular risk.
• Enhance clinical confidence and adoption of automated diagnostic systems.

Main Methods:

• Trained a CNN-RNN model on overnight ECG signals from three pediatric databases (CHAT, CFS, UofC).
• Applied Grad-CAM (XAI) to interpret model decisions and extract relevant ECG features.
• Evaluated diagnostic performance using 4-class Cohen's Kappa.

Main Results:

• SleepECG-Net achieved diagnostic performance (Cohen's Kappa) of 0.410 (CHAT), 0.335 (UofC), and 0.249 (CFS).
• Model performance correlated with OSA severity and cardiovascular risk.
• XAI identified known OSA-related ECG patterns and potential cardiovascular risk indicators (e.g., QT interval variations).

Conclusions:

• SleepECG-Net offers an interpretable approach to diagnosing pediatric OSA severity from ECG.
• The model provides valuable cardiac risk information, potentially improving diagnosis and management.
• This interpretable AI method can increase clinician trust and facilitate the clinical integration of automated systems.