An Interpretable CNN-Transformer for Pediatric Sleep Apnea Diagnosis using ECG

Pediatric obstructive sleep apnea (OSA) is a common respiratory disorder that has been associated with increased cardiovascular risk. The standard diagnosis is polysomnography, but its complexity, cost, and inconvenience lead to underdiagnosis. To address this situation, we propose for the first time a simplified alternative using the overnight electrocardiogram (ECG) and a hybrid model based on a convolutional neural network and a transformer network to estimate the severity of pediatric OSA. In addition, the Gradient-weighted Class Activation Mapping (GradCAM) method is proposed to interpret the model results. For the development of the study, 2,591 recordings from the Childhood Adenotonsillectomy Trial (CHAT) and University of Chicago (UofC) databases were used. The model achieved a 4-class Cohen’s Kappa of 0.392 in CHAT and 0.346 in UofC. GradCAM highlighted bradycardia-tachycardia patterns, and PQ and QT segments, as well as identified U waves. Therefore, this approach may improve the diagnosis of pediatric OSA and provide new related cardiac information, thus encouraging the adoption of automated systems in clinical settings.