0277 Deep learning revealed associations between altered temporal correlations in motor activity and Parkinson's risk. (25th May 2022)

• Record Type: Journal Article
• Authors: Zheng, Xi; Sun, Haoqi; Yang, Jingyun; Yu, Lei; Gao, Lei; Buchman, Aron; Bennett, David; Westover, M Brandon; Hu, Kun; Li, Peng
• Is Part Of: Sleep. Volume 45(2022)Supplement 1
• Page Start: A124
• Page End: A125
• 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/zsac079.275
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Motor activity in healthy young adults displays fractal patterns with similar temporal correlations at different timescales. Altered fractal patterns were observed in patients with Parkinson's disease. This study aimed to determine whether altered fractal patterns also predict the risk of Parkinsonism. Methods: We studied 982 participants (age: 80.12±7.27 [SD]; 750 females) from the Rush Memory and Aging Project, who had at least one actigraphy assessment, had no symptoms of Parkinsonism at actigraphy baseline, and had follow-up clinical assessments. Detrended fluctuation analysis was performed on baseline actigraphy to determine the fractal patterns. Specifically, the activity fluctuation (around the trend) was computed at multiple timescales (n) ranging from 3-600 min. An exponential function with a variable scaling factor α was used to fit the local fluctuation function with respect to n. The α(n) that represented the temporal correlations was fed into a convolutional neural network (CNN) model whose output was further used as the input of a Cox proportional hazards model to predict the time to incident Parkinsonism. Covariates at baseline considered include age, sex, education, cognition, motor function, chronic health assessment, and actigraphy-derived measures including physical activity level, rest-activity/activity-rest transition probabilities, interdaily stability, and intradaily variability. Results: There were 412 subjects who developed parkinsonism (in 4.75±3.13 [SD] years from baseline). Based on the gradient of hazard function (with respect to α) from the CNN model (estimated feature importance), the α in three timescale regions (i.e., 3-5 min, 12-20 min, and 270-600 min) contributed significantly to the prediction. Consistently, in separate Cox models with adjustment of age, sex, and education, the mean α at timescales 3-5 min was inversely associated with incident parkinsonism (for 1-SD increase, hazard ratio [HR]=0.82, 95% CI: 0.78-0.92, p<0.0001); The mean α at timescales 270-600 min was also inversely associated with incident Parkinsonism (HR=0.87, 95% CI: 0.78-0.96, p=0.008); And the mean α at timescales 10-25 min was marginally, positively associated with incident Parkinsonism (HR=1.10, 95% CI: 0.99-1.22, p=0.08). Conclusion: Altered temporal correlations at specific timescales in motor activity predicted the risk of Parkinsonism. Support (If Any): NIH RF1AG064312, RF1AG059867, R01AG017917, R01AG56352, the BrightFocus Foundation A2020886S.