Accurate and Interpretable Deep Learning Model for Sleep Staging in Children with Sleep Apnea from Pulse Oximetry

Identification of sleep stages is crucial in the diagnosis of sleep-related disorders but relies on the labor-intensive and manual scoring of overnight polysomnography (PSG) recordings. To simplify the sleep staging process, deep learning (DL) algorithms have been proposed to automatically analyze pulse rate (PR) and blood oxygen saturation (SpO2) signals from pulse oximetry in children with obstructive sleep apnea (OSA). However, existing approaches are perceived as black boxes, limiting their implementation in clinical settings. Accordingly, we develop a DL architecture based on a U-Net to automatically perform 4-class sleep stage classification (wake, light sleep, deep sleep, and rapid-eye movement sleep) using entire-night PR and SpO2 recordings. Furthermore, Semantic Segmentation via Gradient-Weighted Class Activation Mapping (Seg-Grad-CAM), an eXplainable Artificial Intelligence methodology, is proposed to provide an interpretation of the sleep scoring process. PR and SpO2 from 1,633 PSG recordings obtained from the Childhood Adenotonsillectomy Trial database were used for these purposes. The U-Net model showed a high performance for the 4-stage classification procedure in an independent set, with 78.2% accuracy and 0.696 Cohen’s kappa. The Seg-Grad-CAM heatmaps revealed that the PR signal has a higher contribution than SpO2 towards sleep staging, while also showing the key roles of mean and variance in PR amplitude, along with changes in the content of PR spectral bands, in the sleep staging process. These findings suggest that an explainable DL model to analyze pulse oximetry signals could be integrated in the clinical environment for automatic sleep staging in abbreviated pediatric OSA tests.