Beyond the (Log)book: Comparing Accelerometer Nonwear Detection Techniques in Toddlers

BACKGROUND: Accelerometers are increasingly used to measure physical activity and sedentary time in toddlers. Data cleaning or wear time validation can impact outcomes of interest, particularly in young children who spend less time awake. However, no study has systematically compared wear time validation strategies in toddlers. As such, the objective of this study is to compare different fully automated methods of distinguishing wear and nonwear time (counts and raw data algorithms) methods to the semi automated (counts with logbooks) criterion method in toddlers. METHODS: We recruited 109 toddlers (age 12-35 months) as part of the iPLAY study to wear an ActiGraph w-GT3X-BT accelerometer on the right hip for ~7 consecutive days (removed for sleep and water activities). Parents completed a logbook to indicate monitor removal and nap times. We tested 15 nonwear detection methods grouped into four main categories: semi-automated logbook, consecutive 0 counts, modified consecutive 0 counts (Troiano and Choi) and raw data methods (van Hees and Ahmadi). Using semi-automated logbooks as the criterion standard (all wear and wake-time only wear), we calculated the accuracy and F1 scores (a metric which balances precision and recall) and compared overall wear time with a two one-sided test of equivalence. RESULTS: Participant daily wear time ranged from 556 to 684 min/day depending on method. Accuracy and F1 score ranged from 86% to 95%. Five methods were considered equivalent to the AllWear nonwear criterion (true wear time including sleep-time wear), with only one equivalent to the AwakeWear criterion. Mean absolute differences were lower for the AllWear criterion but ranged from 49 to 192 min/day. CONCLUSIONS: The 5min0count, 10min_0count, 30min_0count, Troiano60s and Ahmadi methods provide high accuracy and equivalency when compared with semi-automated cleaning using logbooks. This paper provides insights and quantitative results that can help researchers decide which method may be the most appropriate given their population of interest, sample size and study protocol.