0435 Autoscoring of Sleep and Associated Events Versus a Reference Scorer Competing with Three Additional Manual Scorings: A Clinical Validation Study. (27th May 2020)

• Record Type: Journal Article
• Authors: Anderer, P; Ross, M; Cerny, A; Moreau, A
• Is Part Of: Sleep. Volume 43(2020)Supplement 1
• Page Start: A167
• Page End: A167
• Publication Date: 2020-05-27
• 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/zsaa056.432
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Manual scoring of polysomnographic (PSG) data is a time-consuming and tedious process with noticeable inter-rater variability. Autoscoring may overcome these limitations if it delivers valid results. The goal of this study was to validate a comprehensive autoscoring system in a clinically representative population. Methods: The validation data consisted of 97 PSGs in patients with sleep-related breathing disorder, scored manually by a reference scorer and three further experts. The Somnolyzer autoscoring system combined pattern recognition for events such as spindles, k-complexes, slow-waves, eye-movements, apneas, hypopneas, desaturations and leg movements with an artificial intelligence classifier consisting of a bidirectional long short-term memory recurrent neural network (RNN) architecture. Intra-class correlation coefficients (ICC) for absolute agreement were determined for the commonly used metrics in sleep medicine to compare both, the three human expert scorings and the autoscoring versus the reference scoring. Results: ICC coefficients for autoscoring and the three manual scorings versus the reference scoring were for sleep efficiency: .95, .83, .91, .93; N1(%): .71, .44, .39, .56; N2(%): .87, .63, .55, .45; N3(%): .80, .62, .44, .32; R(%): .92, .89, .91, .88; arousal index: .88, .81, .22, .78; PLMI: .97, .88, .86, .91; AHI: .91, .89, .87, .78; OA: .94, .89, .91, .90; CA: .96, .96, .96, .82; MA: .93, .77, .43, .41. Thus, the ICCs between autoscoring and the reference scoring were equal or higher than the ICCs between any of the three manual scorings and the reference scoring for all endpoints. Conclusion: Autoscoring of PSGs based on artificial intelligence outperformed even the best of three human expert scorers. Since the autoscoring performs pattern recognition in real-time, the final autoscoring results are available immediately after the end of the recording. Support: All authors are employees of Philips