An Explainable Deep-Learning Approach to Detect Pediatric Sleep Apnea From Single-Channel Airflow.

Area of Science:

• Artificial Intelligence in Medicine
• Pediatric Sleep Medicine
• Biomedical Signal Processing

Background:

• Single-channel airflow analysis shows promise for diagnosing pediatric obstructive sleep apnea (OSA).
• Traditional methods using feature engineering limit complex respiratory pattern identification and automated diagnostic performance.
• Deep learning and explainable AI (XAI) offer potential for improved accuracy and transparency in OSA diagnosis.

Purpose of the Study:

• To develop and validate a deep learning model for estimating pediatric OSA severity using single-channel airflow.
• To ensure transparency in automated OSA diagnosis through explainable AI techniques.
• To compare the effectiveness of Grad-CAM and SHAP for identifying key airflow features in OSA detection.

Main Methods:

• Utilized 3,672 overnight airflow recordings from four pediatric datasets.
• Trained a convolutional neural network (CNN) regression model to estimate the apnea-hypopnea index (AHI) and predict OSA severity.
• Employed Gradient-Weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• The CNN model achieved high concordance between estimated and actual AHI (ICC 0.69–0.87).
• Demonstrated high diagnostic performance for three OSA severity cutoffs (accuracies 82.03%–99.03%).
• Interpretability analysis confirmed the CNN's accurate identification of apneic events, with Grad-CAM and SHAP providing complementary insights.

Conclusions:

• The developed interpretable deep learning tool accurately detects pediatric OSA from airflow data.
• This approach facilitates early, objective diagnosis and supports clinical decision-making.
• The explainable AI enhances model credibility and usability, paving the way for clinical translation.