Estimating Sleep Duration based on Non-wear Time Imputation and Sleep Algorithms Using Accelerometer Data from the Korea National Health and Nutrition Examination Survey

Objectives: This study aimed to improve the accuracy of sleep duration estimation from accelerometer data by addressing missing values caused by daytime non-wear periods. It also explored how different imputation methods and sleep algorithms affect sleep duration estimates.Methods: We used accelerometer data from the Korea National Health and Nutrition Examination Survey (KNHANES). The study population included adults aged 19–64 years who wore accelerometers for at least 10 hours per day between 9 a.m. and 9 p.m. on at least three days. Non-wearing periods were defined as intervals of 60 or more consecutive minutes with zero activity counts. For missing value imputation, we applied zero-inflated Poisson log-normal (ZIPLN) mixed model imputation, multivariate imputation by chained equations (MICE), and a method combining predictive mean matching (PMM) with each method (ZIPLN+PMM, MICE+PMM). After imputation, sleep duration was estimated using six combinations of the Sadeh and Cole–Kripke algorithms with and without the Tudor–Locke approach.Results: Cole–Kripke consistently produced longer sleep durations than the Sadeh algorithm. The raw data overestimated sleep duration relative to the imputed datasets, likely due to the misclassification of daytime non-wearing periods as sleep. Applying the Tudor– Locke 60-minute criterion resulted in shorter and more conservative sleep estimates. These trends were consistent in the complete wear group analyses.Conclusions: Imputation methods can reduce overestimation of sleep caused by daytime non-wear. ZIPLN approaches better preserve the distributional structure of accelerometer data. Cole-Kripke estimated longer sleep than Sadeh, and Tudor-Locke yielded shorter durations with longer time windows.