|
Abstract:
|
Accelerometers are used frequently in biomedical research including in the clinical trial setting to derive key endpoints to understand physical activity, sedentary behavior, sleep and their relationships with many cardiovascular health outcomes. They are now incorporated into various “mobile” and “wearable” products and yield massive amounts of data. Despite the size of the data generated, classical statistical principles regarding missing data need to be considered. Specifically, a critical processing step involves identifying which data points, if any, reflect non-wear or missing values. We discuss an ensemble learner for identifying non-wear periods and multiple imputation (MI) as a potential tool for handling missingness in the analysis once such periods are identified. Performance of the ensemble learner demonstrates the strength of such an approach in the free living context when true labels are not available. Considerable heterogeneity in statistical properties is demonstrated among MI strategies where special considerations need to be made that acknowledge the time-series and tri-variate nature of the data. Guidelines for application of MI to this setting are provided.
|
