Deepsleep: Fast and Accurate Delineation of Sleep Arousals at Millisecond Resolution by Deep Learning

Background: Sleep arousals are transient periods of wakefulness punctuated into sleep. Excessive sleep arousals are associated with many negative effects including daytime sleepiness and sleep disorders. High-quality annotation of polysomnographic recordings is crucial for the diagnosis of sleep arousal disorders. Currently, sleep arousals are mainly annotated by human experts through looking at millions of data points manually, which requires considerable time and effort. Methods: We used the polysomnograms of 2,994 individuals from two independent datasets (i) PhysioNet Challenge dataset (n=994), and (ii) Sleep Heart Health Study dataset (n=2000) for model training (60%), validation (15%), and testing (25%). We developed a deep convolutional neural network approach, DeepSleep, to automatically segment sleep arousal events. Our method captured the long-range and short-range interactions among physiological signals at multiple time scales to empower the detection of sleep arousals. A novel augmentation strategy by randomly swapping similar physiological channels was further applied to improve the prediction accuracy. Findings: Compared with other computational methods in sleep study, DeepSleep features accurate (area under receiver operating characteristic curve of 0.93 and area under the precision recall curve of 0.55), high-resolution (5-millisecond resolution), and fast (10 seconds per sleep record) delineation of sleep arousals. This method ranked first in segmenting non-apenic arousals when evaluated on a large held-out dataset (n=989) in the 2018 PhysioNet Challenge. We found that DeepSleep provided more detailed delineations than humans, especially at the low-confident boundary regions between arousal and non-arousal events. This indicates that in silico annotations is a complement to human annotations and potentially advances the current binary label system and scoring criteria for sleep arousals. Interpretation: The proposed deep learning model achieved state-of-the-art performance in detection of sleep arousals. By introducing the probability of annotation confidence, this model would provide more accurate information for the diagnosis of sleep disorders and the evaluation of sleep quality. Funding Statement: This work is supported by NSF-US14-PAF07599 (CAREER: On-line Service for Predicting Protein Phosphorylation Dynamics Under Unseen Perturbations NSF), AWD007950 (Digital Biomarkers in Voices for Parkinson's Disease American Parkinson's Disease Association), University of Michigan O'Brien Kidney Translational Core Center, 19AMTG34850176 (American Heart Association and Amazon Web Services3.0 Data Grant Portfolio: Artificial Intelligence and Machine Learning Training Grants), and Michael J. Fox Foundation #17373. Declaration of Interests: YG receives personal payment from Eli Lilly and Company, Genentech Inc, F. Hoffmann-La Roche AG, and Cleerly Inc; holds equity shares at Cleerly Inc and Ann Arbor Algorithms Inc; receives research support from Merck KGaA as research contracts and Ryss Tech as unrestricted donation. Ethics Approval Statement: Not required.