Combined explainable deep learning model to predict pediatric sleep apnea from ECG and SpO₂.

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Pediatric Sleep Medicine

Background:

• Pediatric obstructive sleep apnea (OSA) diagnosis can be complex.
• Deep learning (DL) and eXplainable Artificial Intelligence (XAI) offer potential for improved diagnostic models.
• Concurrent analysis of electrocardiogram (ECG) and oxygen saturation (SpO2) for pediatric OSA has not been previously explored.

Purpose of the Study:

• To develop and validate an explainable deep learning model for pediatric OSA identification.
• To concurrently analyze overnight SpO2 and ECG signals for OSA diagnosis.
• To elucidate the relationship between respiratory events and cardiac patterns in pediatric OSA using XAI.

Main Methods:

• An explainable DL approach integrating convolutional neural networks with overnight SpO2 and ECG signals was developed.
• The SHapley Additive exPlanations (SHAP) XAI technique was employed to interpret model decisions.
• A total of 3,320 pediatric patients from CHAT, PATS, and UofC databases were analyzed.

Main Results:

• The model achieved Cohen's kappa scores of 0.549 (CHAT), 0.457 (PATS), and 0.378 (UofC).
• Shapley values indicated SpO2 is crucial for moderate-severe OSA, while ECG is important for mild/no OSA, demonstrating complementarity.
• SHAP analysis revealed correlations between SpO2 desaturations, apneic events, and cardiac changes (e.g., bradycardia-tachycardia, wave variations).

Conclusions:

• The interpretable DL approach effectively integrates respiratory and cardiac data for pediatric OSA diagnosis.
• The model's ability to explain decisions by highlighting relevant SpO2 and ECG patterns supports clinical adoption.
• This method offers a promising tool to enhance the accuracy and understanding of pediatric OSA diagnosis.