0850 Self-report And Polysomnography Sleep And Mortality In Adults: A Machine Learning Replication Analysis. (27th May 2020)

• Record Type: Journal Article
• Authors: Wallace, M L; Peppard, P; Coleman, T L; Mentch, L; Buysse, D J; Hall, M H; Redline, S; Hagen, E H
• Is Part Of: Sleep. Volume 43(2020)Supplement 1
• Page Start: A323
• Page End: A324
• 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/zsaa056.846
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Individual sleep health characteristics (e.g. efficiency, timing, duration, architecture) and signs and symptoms of sleep disorders (e.g., difficulty falling and staying asleep, apnea hypopnea index, measures of oxygen desaturation) predict mortality in adults using traditional regression methods. However, it is important to examine and compare their predictive abilities in context of other established non-sleep predictors using high-dimensional methods that better reflect the complexity of the data. Therefore, we applied a novel random forest machine learning (RFML) hypothesis-testing framework to data from the Sleep Heart Health Study (SHHS) and the Wisconsin Sleep Cohort (WSC) to determine which risk factor domains (sleep, physical health, sociodemographic factors, medications, health behaviors, mental health) and sleep subdomains (self-report and polysomnography sleep health characteristics and signs and symptoms of sleep disorders) predict time to mortality in adults. Methods: We harmonized 82 predictors across SHHS and WSC (32 sleep, 24 physical health, 8 sociodemographic, 9 medications, 4 mental health, 5 health behaviors) and fit sociodemographic-adjusted and fully-adjusted RFML models in each cohort to test the overall predictive importance of each domain and sleep subdomain. Permutation-based p-values and unbiased variable importance metrics (change in Harrell's C *100, ΔC) were computed and summarized with medians across 20 independent subsampled testing sets in each cohort. Results: In the fully-adjusted SHHS and WSC models, the most predictive domains were physical health (SHHS p<0.001, ΔC=1.48; WSC p=0.002, ΔC=2.68) and sleep (SHHS p=0.008, ΔC=0.71; WSC p=0.044, ΔC=1.65). Sleep subdomains were not significant in the fully adjusted model. However, the sociodemographic-adjusted models indicated that the predictive importance of sleep may be driven by polysomnography sleep health characteristics in SHHS (p=0.026, ΔC =0.77) and polysomnography signs of sleep apnea in WSC (p<0.001, ΔC=3.20). Conclusion: Sleep is a strong predictor of mortality in adults that should be considered among other more routinely used predictors. Future research should examine differences and similarities between SHHS and WSC that may explain the finding that different aspects of sleep were important in each cohort. Support: NIA grant AG056331, NHLBI grant HL114473, NHLBI grant R01HL62252, NIA grant R01AG036838, NIA grant R01AG058680.