0315 Representation Of Polysomnography Recordings As Low Dimensional Trajectories. (12th April 2019)

• Record Type: Journal Article
• Authors: Solelhac, Geoffroy; Brigham, Marco; Bouchequet, Paul; Andrillon, Thomas; Chennaoui, Mounir; Pennec, Erwan Le; Rey, Marc; Leger, Damien
• Is Part Of: Sleep. Volume 42(2019)Supplement 1
• Page Start: A129
• Page End: A129
• Publication Date: 2019-04-12
• 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/zsz067.314
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Polysomnography (PSG) recording is the gold standard in the study of sleep. Qualitative evaluation of sleep (sleep scoring) is based on the visual identification of sleep stages by human experts according to the traditional Rechtschaffen and Kales standard and the more recent American Academy of Sleep Medicine standard. Considering that PSG is the best system for recording a sleep session, current rules show some limitations, such as inter-scorer accuracy (~80%), ambiguous epochs or paradoxical insomnia. Dimensional reduction methods allow us to visualize complex and synchronized data in two or three dimensions as PSG requires more precise and comprehensive analysis. We have chosen here to compare different dimension reduction methods to visualise PSG. The aim of this study is to visually find known parameters of sleep with low dimension representations of polysomnography. Methods: We use 2 types of dimensional reduction models to project polysomnography recordings into 2 or 3 dimensions trajectories: Firstly, a Principal Component Analysis model that reduces the number of variables and makes informations less redundant. Secondly, an unsupervised machine learning model (autoencoder) that learns to compress information into lower-dimension representation (latent space). Importantly, this compression forces the model to represent similar signals into the same regions of the latent space. We applied these algorithms on a dataset of 10 chronic insomniacs, 10 patients with sleep state misperception and 10 good sleepers. Visual criterions selected were the slow wave gradient, the REM sleep differentiation and the visualization of sleep stages. Results: With these two models we find known parameters of the sleep independently of the sleep scoring. These parameters seem invariable for each patient. Conclusion: These low-dimensional representations of polysomnography make it possible to visually find known sleep parameters with interesting visual invariability on a dataset of 30 subjects including insomiac patients and sleep state misperception. These representions are promising tools to complement the standard hypnogram and may enable better identification of sleep disorders. Support (If Any): Banque publique d'investissement, Dreemcare Project.