An explainable deep-learning architecture for pediatric sleep apnea identification from overnight airflow and oximetry signals

Abstract

Deep-learning algorithms have been proposed to analyze overnight airflow (AF) and oximetry (SpO2) signals to simplify the diagnosis of pediatric obstructive sleep apnea (OSA), but current algorithms are hardly interpretable. Explainable artificial intelligence (XAI) algorithms can clarify the models-derived predictions on these signals, enhancing their diagnostic trustworthiness. Here, we assess an explainable architecture that combines convolutional and recurrent neural networks (CNN + RNN) to detect pediatric OSA and its severity. AF and SpO2 were obtained from the Childhood Adenotonsillectomy Trial (CHAT) public database (n = 1,638) and a proprietary database (n = 974). These signals were arranged in 30-min segments and processed by the CNN + RNN architecture to derive the number of apneic events per segment. The apnea-hypopnea index (AHI) was computed from the CNN + RNN-derived estimates and grouped into four OSA severity levels. The Gradient-weighted Class Activation Mapping (Grad-CAM) XAI algorithm was used to identify and interpret novel OSA-related patterns of interest. The AHI regression reached very high agreement (intraclass correlation coefficient > 0.9), while OSA severity classification achieved 4-class accuracies 74.51% and 62.31%, and 4-class Cohen’s Kappa 0.6231 and 0.4495, in CHAT and the private datasets, respectively. All diagnostic accuracies on increasing AHI cutoffs (1, 5 and 10 events/h) surpassed 84%. The Grad-CAM heatmaps revealed that the model focuses on sudden AF cessations and SpO2 drops to detect apneas and hypopneas with desaturations, and often discards patterns of hypopneas linked to arousals. Therefore, an interpretable CNN + RNN model to analyze AF and SpO2 can be helpful as a diagnostic alternative in symptomatic children at risk of OSA.