0553 At-home Detection of REM Sleep Behavior Disorder using a Machine Learning Approach and Wrist Actigraphy. (25th May 2022)

• Record Type: Journal Article
• Authors: Brink-Kjaer, Andreas; Gupta, Niraj; Marin, Eric; Zitser, Jenny; Sum-Ping, Oliver; Hekmat, Anahid; Bueno, Flavia; Cahuas, Ana; Jennum, Poul; Sorensen, Helge; Mignot, Emmanuel; During, Emmanuel
• Is Part Of: Sleep. Volume 45(2022)Supplement 1
• Page Start: A243
• Page End: A244
• Publication Date: 2022-05-25
• 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/zsac079.550
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Isolated rapid-eye-movement (REM) sleep behavior disorder (iRBD) affects over 1% of middle-aged and older adults and is in most cases a prodromal stage of alpha-synucleinopathy. However, a small fraction of them is currently diagnosed due to poor access to the gold-standard diagnostic procedure polysomnography (PSG). We aimed to test an ambulatory diagnostic procedure for iRBD based on wrist actigraphy alone and combined with a short questionnaire on nonmotor symptoms. Methods: A total of 35 PSG-confirmed iRBD and 28 age-matched clinic and community control participants with and without a sleep disorder (1:1 ratio) wore high-frequency (25 Hz) wrist actigraphy for at least 7 nights and completed sleep diaries. Raw accelerometer data recorded during sleep was analyzed by deriving an activity count and extracting movement-related features for each night. Additionally, participants completed the Innsbruck RBD inventory (RBD-I) and a 3-item questionnaire on hyposmia, constipation, and orthostasis. We fitted machine learning models, specifically, boosted decision trees, in a leave-one-out cross-validation framework to classify iRBD patients from controls based on either actigraphy or questionnaire data. For each participant, model predictions from actigraphy were averaged across all available nights. Results: The boosted decision trees classified iRBD with an area under the receiver-operator-characteristics (ROC) curve (AUC) of 0.972, a sensitivity of 97.1%, and a specificity of 89.3%. Analyses revealed that performance plateaued after one week of actigraphy. Best single feature "short immobile bursts" achieved an AUC of 0.958, a sensitivity of 94.3%, and a specificity of 78.6%. In this population, RBD-I item 3 best discriminated between groups with an AUC of 0.892, a sensitivity of 91.4%, and a specificity of 85.7%. The combination of a positive RBD-I item 3 and a positive actigraphy-based classification achieved a sensitivity of 88.6% and a specificity of 96.4%. Conclusion: High-frequency actigraphy using machine learning detects iRBD with high accuracy. Addition of a single RBD question to this procedure increased specificity. These results need to be validated in a larger sample and lay the groundwork for an ambulatory screening paradigm in the general population. Support (If Any): The Klarman Family Foundation and the Feldman Foundation Ca.