0142 Automatic Detection of Cortical Arousals in Sleep using Bi-direction LSTM Networks. (27th April 2018)

• Record Type: Journal Article
• Authors: Brink-Kjær, A; Olesen, A N; Jespersen, C A; Peppard, P E; Jennum, P J; Sørensen, H B; Mignot, E
• Is Part Of: Sleep. Volume 41(2018)Supplement 1
• Page Start: A55
• Page End: A56
• Publication Date: 2018-04-27
• Subjects: Sleep -- Physiological aspects -- Periodicals; Sleep disorders -- Periodicals; Sommeil -- Aspect physiologique -- Périodiques; Sommeil, Troubles du -- Périodiques; Sleep disorders; Sleep -- Physiological aspects; Sleep -- physiological aspects; Sleep Wake Disorders; Psychophysiology; Electronic journals; Periodicals; 616.8498
• DOI: 10.1093/sleep/zsy061.141
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Cortical arousals are transient events that occur during sleep. Although they can occur naturally, arousals are often used to evaluate sleep-wake dysfunction. The gold standard for detecting arousals is visual inspection of polysomnography recordings. Manual annotation of arousals is time consuming and has been shown to have a high inter- and intra-scorer variation. This study proposes a method to fully automate detection of arousals using recent advances in machine learning. Methods: The proposed method in this study extracted features from electroencephalography (EEG), electrooculography (EOG) and chin electromyography (EMG) to compute a probability of arousals through a bi-directional long short-term memory neural network. The study used a dataset of 233 nocturnal PSGs of population-based samples from Wisconsin Sleep Cohort (WSC) and 30 nocturnal PSGs of clinical samples from the Stanford Sleep Cohort (SSC). The model was trained on 186 recordings from WSC and annotations from two scorers. The model was tested on 47 recordings from WSC and then compared to a set of 3 annotations from 9 independent scorers on 30 recordings from both cohorts by measure of Fleiss' Kappa (level of agreement greater than chance). Results: The model obtained a precision of 0.79, a recall of 0.8 and F1-score of 0.79 on the 47 recordings from WSC. The model was robust to different sleep stages showing an F1-score of 0.71 ± 0.19, 0.8 ± 0.13, 0.89 ± 0.18 and 0.8 ± 0.17 (mean ± SD) for N1, N2, N3 and REM sleep, respectively. Preliminary results comparing the scorers show a Fleiss' Kappa of 0.38 ± 0.12, while including the model predictions result in a Fleiss' Kappa of 0.4 ± 0.1. Conclusion: Cortical arousals were detected automatically with the proposed algorithm with a high performance and robustness to different sleep stages. Preliminary results comparing nine independent scorers demonstrates a low inter-scorer reliability with a similar agreement to the model predictions. Support (If Any): Klarman Family Foundation, grants from H. Lundbeck A/S, the Lundbeck Foundation, the Technical University of Denmark, and the Center for Healthy Aging, University of Copenhagen.