A convolutional neural network to classify sleep stages in pediatric sleep apnea from pulse oximetry signals

Characterization of the sleep and wake stages is essential in the diagnosis of pediatric obstructive sleep apnea (OSA). The onerous requirements and limitations of overnight polysomnography (PSG), the gold standard, have led to the search for simplified sleep scoring systems. Accordingly, the aim of this study was to assess the usefulness of a convolutional neural network (CNN)-based deep-learning architecture fed with pulse oximetry signals to automatically classify sleep stages in symptomatic children at risk of OSA. Nocturnal pulse rate (PR) and blood oxygen saturation (SpO2) from 429 pediatric OSA patients were employed for this purpose. A 2D CNN architecture was trained using 30-s epochs from PR and SpO2 signals for the automatic classification of the three main sleep stages: wake (W), non-Rapid Eye Movement (non-REM), and REM sleep. The proposed 2D CNN model showed a promising diagnostic performance for the three-stage classification procedure (W/NREM/REM) in an independent test set, with 83.1% accuracy and 0.680 Cohen's kappa, outperforming 1D CNN models trained using PR or SpO2 signals alone. Furthermore, the total sleep time calculated for each subject using the 2D CNN model showed high agreement with the manually scored from PSG (intra-class correlation coefficient of 0.677). These results were superior to previous studies focused on the automated detection of sleep stages in pediatric OSA patients using photoplethysmography and PR signals derived from pulse oximetry. Therefore, joint analysis of PR and SpO2 signals using CNNs can be helpful detect sleep stages in at-home pulse oximetry tests for pediatric OSA diagnosis.