A Novel Method To Estimate Free-living Physical Behaviors From A Wrist-Worn ActiGraphTMAccelerometer

Abstract

BACKGROUND: Despite the proliferation of body-worn sensors to assess physical behaviors (e.g. physical activity and sedentary behavior; PBs), processing accelerometer data collected from free-living (FL) individuals remains a challenge. PURPOSE: To develop a method to estimate FL PBs from wrist-worn ActiGraph accelerometer count data. METHODS: Forty-nine participants (X +/- SD age:20.4±1.3 yrs, 45.8% male) wore an ActiGraph wGT3X-BT on their non-dominant wrist during four, 1-hour FL sessions. Sessions were video-recorded and coded using a direct observation (DO) system that provided criterion measures, including intensity category. All steps in the novel method were developed using a subset (n=44) of participants and tested on the remaining (n=5). The first step identifies inactive bouts (e.g. sedentary and standing behaviors) using a vector-magnitude threshold, chosen via a grid-search to maximize the positive predictive value (PPV) of inactive classification. The second step uses statistical features of the count data and bout length (seconds) as inputs to random forest models to estimate active (light, moderate, vigorous) vs inactive periods. Percent agreement between criterion-measured (DO) and method estimated PB was evaluated using second-by-second data. Gini impurity index was used to measure the importance of statistical features used as inputs to the random forest models. RESULTS: The first step identified a vector magnitude threshold that resulted in a PPV 95.71% for inactive classification. In the holdout sample, steps 1 and 2 correctly classified 84.1% of inactive and active bouts. Gini impurity index revealed that bout length (seconds) was most important among statistical features. CONCLUSIONS: The new method performed well at estimating FL PBs from wrist-worn ActiGraph accelerometer counts. Future work should investigate how to improve method estimates using PB bout duration and other novel accelerometer statistical features.Supported by NIH NIDDK 1R01DK110148.