Sleep Stage Classification of Pediatric Patients with Sleep-Disordered Breathing using Airflow Signals.

Sleep staging is a critical task for sleep health assessment. Recent work has shown that breathing variability may be characteristic of particular sleep stages. In this work, we train machine learning models on sleep airflow signals from pediatric patients to assess how well sleep stage can be classified from airflow signals, particularly in subjects with sleep-disordered breathing. We perform a 3-class classification task (Wake/NREM/REM) and achieve 66.9% balanced accuracy (BA) when classifying the sleep stage of healthy subjects, 61.4% BA when classifying the sleep stage of subjects with mild AHI (1 ≤ AHI < 5), 58.0% BA when classifying the sleep stage of subjects with moderate AHI (5 ≤ AHI ≤ 10), and 48.9% BA when classifying the sleep stage of subjects with high AHI (10< AHI). We further show that our model is able to achieve > 60% BA when classifying the sleep stage in which apnea and hypopnea events occur for subjects with AHI < 5. This indicates that although this approach is affected by apnea and hypopnea episodes, it is robust enough to work for sleep apnea subjects. These results show the promise of airflow signals as a surrogate signal to perform sleep staging when EEG signals are not available.Clinical relevanceThis study establishes the potential of airflow signals to assist with sleep staging tasks in both healthy pediatric subjects and subjects with sleep-disordered breathing.