A Pilot Study on the Impact of Accelerometer Data Reduction Algorithm Selection and the Potential Implications on Behavior Change Recommendations (OR08-03-19)

Abstract

Objectives Physical activity (PA) estimates obtained from recent accelerometer data reduction algorithms have not been compared in women of reproductive-age, a population more likely to engage in unstructured and intermittent PA (such as household cleaning, walking) than men. We investigated whether the accelerometer data from the Crouter, Sasaki and Santos-Lozano algorithms: 1) reported significantly different PA estimates; 2) interacted with weight and age to modify PA estimates; and 3) provided different prevalence of adults meeting PA guidelines. Methods At least four days of accelerometer data were collected from 29 women, ages 18 to 38 years, and processed through three algorithms using an Excel model that automatically removed non-wear data and simultaneously calculated PA estimates [i.e., wear minutes, metabolic equivalent minutes (MET-min)]. A combination of mixed-effects linear regression models and bivariate correlation analyses were used to examine associations between accelerometer data with weight, age, and clinical markers of metabolic status across algorithms. Results The Crouter algorithm estimated significantly more wear minutes in Moderate intensity compared to the Sasaki and Santos-Lozano algorithms [+384 (SE 33) and+356 (SE 33) minutes, respectively]. There were significant interactions between Crouter estimates of Sedentary/Light and Moderate wear minutes with weight and age (all Pinteraction ≤ 0.001, Santos-Lozano algorithm as the reference). Algorithm selection also provided inconsistent findings in the prevalence of adults meeting PA guidelines. Conclusions Recently proposed data reduction algorithms varied in their PA estimates in women of reproductive age. Algorithm selection interacted with weight and age to influence PA estimates and provided inconsistent classification of those who met PA guidelines. Thus, depending on the algorithm selected, behavior change recommendations might differ for each individual due to varying PA estimations. Larger sample sizes are needed to confirm these findings. Funding Sources This research is partially supported by the Cornell University Human Ecology Alumni Association. The first author is currently being supported by the National Institutes of Health.