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

OBJECTIVE

Approaches based on a single-channel airflow has shown great potential for simplifying pediatric obstructive sleep apnea (OSA) diagnosis. However, analysis has been limited to feature-engineering techniques, restricting identification of complex respiratory patterns, and reducing diagnostic performance in automated models. Here, we propose deep-learning and explainable artificial intelligence (XAI) to estimate the pediatric OSA severity from airflow, while ensuring transparency in automatic decisions. Technology or Method: We used 3,672 overnight airflow recordings from four pediatric datasets. A convolutional neural network (CNN)-based regression model was trained to estimate the apnea-hypopnea index (AHI) and predict OSA severity. We evaluated and compared Gradient-Weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) to identify the airflow regions where the CNN focuses for predictions.

RESULTS

The proposed model demonstrated high concordance between the actual and estimated AHI (intraclass correlation coefficient from 0.69 to 0.87 in the test group), and high diagnostic performance: four-class Cohen's kappa between 0.37 and 0.43 and accuracies of 82.03%, 97.09%, and 99.03% for three OSA severity cutoffs (i.e. 1, 5, and 10 e/h) in the test group. The interpretability analysis with Grad-CAM and SHAP revealed that the CNN accurately identifies apneic events by focusing on their onset and offset. Both techniques provided complementary information about the model's decision-making. While Grad-CAM highlighted respiratory events with abrupt signal changes, SHAP captured more subtle patterns with noise included.

CONCLUSIONS

Accordingly, our model can help automatically detect pediatric OSA and offers clinicians an explainable approach that enhances credibility and usability, thus providing a path toward clinical translation in early diagnosis.

CLINICAL IMPACT: This study presents an interpretable deep-learning tool using airflow to accurately detect pediatric obstructive sleep apnea, enabling early, objective diagnosis and supporting clinical decision-making through identification of relevant respiratory patterns.