An explainable deep learning approach for sleep staging in sleep apnea patients across all age subgroups from pulse oximetry signals

Abstract

Deep-learning (DL) approaches have been developed using pulse rate (PR) and blood oxygen saturation (SpO2) recordings from pulse oximetry to streamline sleep staging, particularly for obstructive sleep apnea (OSA) pa- tients. However, lack of interpretability and validation across patients from a wide range of ages (children, adolescents, adults, and elderly OSA individuals) are two major concerns. In this study, a DL model based on the U-Net framework (POxi-SleepNet) was tailored to accurately perform 4-class sleep staging (wake, light sleep, deep sleep, and rapid-eye movement sleep) in OSA patients across all age subgroups using PR and SpO2 signals. An explainable artificial intelligence (XAI) methodology based on semantic segmentation via gradient-weighted class activation mapping (Seg-Grad-CAM) was also applied to quantitatively interpret the time and frequency characteristics of pulse oximetry recordings that influence sleep stage classification. Overnight PR and SpO2 signals from 17303 sleep studies from six datasets encompassing children, adolescents, adults, and elderly OSA individuals were used. POxi-SleepNet showed high performance for sleep staging in the six databases, with accuracies between 81.5 % and 84.5 % and Cohen’s kappa values from 0.726 to 0.779. It also demonstrated greater generalizability than previous studies. XAI analysis showed the key contributions of mean and variability in PR and SpO2 amplitude, as well as changes in their spectral content across specific frequency bands (0.004–0.020 Hz, 0.020–0.100 Hz, and 0.180–0.400 Hz), for sleep stage classification. These findings indicate that POxi-SleepNet could effectively automate sleep staging and assist in diagnosing OSA across all age groups in clinical settings.