Using a deep learning–based model to characterize sleep staging, researchers were able to accurately and automatically detect patients with and without obstructive sleep apnea (OSA), but accuracy decreased with increasing OSA severity, according to study findings.
Using a deep learning—based model to characterize sleep staging, researchers were able to accurately and automatically detect patients with and without obstructive sleep apnea (OSA); however, accuracy decreased with increasing OSA severity, according to study findings published in the IEEE Journal of Biomedical and Health Informatics.
OSA is among the most common sleep disorders, affecting an estimated 38% of the general population. To diagnose this condition, sleep staging is conducted to assess sleep characteristics and accurately determine the total sleep time. As study authors note, “accurate determination of total sleep time is of paramount importance as it significantly affects the parameters used to assess the severity of OSA.”
Current sleep staging criteria involve 5 different sleep classifications: wake, rapid eye movement (REM) sleep, and 3 stages of non-REM sleep. These stages are classified manually for 30-second epochs of sleep using electroencephalography (EEG), electrooculogram (EOG), and submental electromyogram signals measured during polysomnography.
Study authors note that the overall measurement protocol and sleep-staging process can be time-consuming and laborious, requiring the use of experienced professionals. Moreover, the overall reliability of manual sleep staging may decrease if an individual is experiencing medical conditions, such as OSA, for which classification reliability is worse than that of healthy individuals.
Researchers aimed to develop an accurate deep learning approach to ameliorate the sleep staging process, an innovation that has shown efficacy in enhancing detection of data relevant to patient risk in pediatric settings and breast cancer. Using overnight polysomnographic recordings from a public data set of healthy individuals, Physionet Sleep-EDF (n = 153), and a clinical data set of patients with suspected OSA (n = 891), researchers developed a combined convolutional and long short-term memory neural network to compare classification efficacy, with additional assessments on whether severity of OSA affected accuracy.
Among healthy individuals, the model achieved sleep staging accuracy of 83.7% (κ = 0.77) with a single frontal EEG channel and 83.9% (κ = 0.78) when supplemented with EOG. The model achieved similar accuracy in patients with suspected OSA of 82.9% (single EEG channel) and 83.8% (EEG and EOG channels). The difference of adding an EOG channel was not found to significantly increase the accuracy of the model. When tested on individuals with more severe OSA, single-channel accuracy suffered (76.5%) compared with those without OSA diagnosis (84.5%), indicating the effect of OSA severity on sleep staging efficacy.
“Deep learning enables automatic sleep staging for suspected OSA patients with high accuracy and expectedly, the accuracy lowered with increasing OSA severity,” wrote the study authors. Compared with previously published state-of-the-art methods, researchers highlight that accuracies achieved on sleep staging in the public data set were superior, indicating the potential of a single-channel—based model to enable easy, accurate, and cost-efficient integration of EEG recording into diagnostic ambulatory recordings.
Korkalainen H, Aakko J, Nikkonen S, et al. Accurate deep learning-based sleep staging in a clinical population with suspected obstructive sleep apnea [published online December 19, 2019]. IEEE J Biomed Health Inform. doi: 10.1109/JBHI.2019.2951346.