Explainable hybrid convolutional and transformer network for pediatric sleep apnea diagnosis using nocturnal oximetry.

Pediatric obstructive sleep apnea (OSA) is a breathing disorder marked by pauses in airflow (apneas) and reduced airflow (hypopneas), contributing to neurocognitive and behavioral impairments, cardiovascular complications, and other health issues in affected children. Polysomnography is the gold standard for diagnosis, but complexity, high cost, and limited accessibility issues often lead to underdiagnosis. To address these challenges, we propose a simplified diagnostic approach based on blood oxygen saturation (SpO2) recordings from nocturnal oximetry. A total of 1,609 SpO2 recordings from the Childhood Adenotonsillectomy Trial (CHAT) were analyzed. We developed an interpretable approach leveraging a convolutional-transformer network to estimate pediatric OSA severity. Furthermore, we evaluated the explainable artificial intelligence method Gradient-weighted Class Activation Mapping (Grad-CAM). The model achieved 4-class Cohen's kappa and accuracy of 0.529 and 68.56% in the test set, respectively. The proposed model demonstrated enhanced performance correlating with increasing disease severity, with accuracy values ranging from 82% to 95% at different severity cut-offs, thereby signaling improved diagnostic performance when compared to previous approaches. Furthermore, Grad-CAM identified key SpO2 patterns linked to OSA, such as SpO2 desaturations related to clusters of apneic events and desaturations occurring independently of events. This innovative approach represents a promising alternative for diagnosing OSA and provides valuable insights into respiratory abnormalities associated with pediatric OSA.Clinical Relevance-This study highlights the potential of an interpretable deep-learning approach using overnight oximetry for diagnosing pediatric obstructive sleep apnea. It effectively identifies clinically relevant desaturation patterns associated with the disease and supports its early, objective, and efficient detection in clinical practice.