AN EXPLAINABLE DEEP-LEARNING APPROACH TO DETECT PEDIATRIC SLEEP APNEA FROM SINGLE-CHANNEL AIRFLOW

Approaches based on single-channel airflow show great potential for simplifying pediatric obstructive sleep apnea (OSA) diagnosis, but most analyses have relied on feature-engineering, limiting recognition of complex respiratory patterns and lowering performance in automated models. We propose deep learning and explainable artificial intelligence to estimate pediatric OSA severity from airflow while ensuring transparent predictions. We used 3,672 recordings from four datasets. A convolutional neural network (CNN)-based regression model was trained to estimate the apnea-hypopnea index (AHI) and predict severity. We evaluated Gradient-Weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) to identify airflow regions guiding CNN predictions. The model showed strong concordance between actual and estimated AHI (intraclass correlation coefficient 0.69–0.87 in the test group) and high diagnostic performance: four-class Cohen's kappa 0.37–0.43 and accuracies of 82.03%, 97.09%, and 99.03% for 1, 5, and 10 events/h in the test group. Interpretability analysis revealed that the CNN accurately identified apneic events by focusing on their onset and offset. While Grad-CAM highlighted abrupt signal changes, SHAP captured subtler, noisier patterns. Accordingly, our model enables automatic pediatric OSA detection and provides clinicians with explainable outputs that enhance credibility and usability, paving the way for clinical translation in early diagnosis.