Machine Learning Approaches to Calibrate Wrist-Worn Accelerometry for Physical Activity Assessment in Preschoolers

Physical activity (PA) level is associated with multiple health benefits during early childhood. However, inconsistency remains in the quantification of PA levels in preschoolers. The present study aimed at using machine learning (ML) approaches to develop PA intensity cut-points for wrist-worn accelerometry to assess PA in preschoolers. Wrist- and hip-worn accelerometer data were collected from 34 preschoolers aged 3-5 years old enrolled in childcare. The preschoolers wore the accelerometers simultaneously on their wrists and hips for three consecutive school days. Data was downloaded at 15-second epoch. Two supervised ML models, receiver operating characteristic curve (ROC) and ordinal logistic regression (OLR), and one unsupervised ML model, k-means cluster analysis, were applied to establish wrist-worn accelerometer vector magnitude (VM) cut-points to classify PA levels into sedentary behavior (SED), light PA (LPA), moderate PA (MPA), and vigorous PA (VPA). Using PA intensity levels identified by the Butte et al. (2014b), hip-worn accelerometer VM cut-points were used as the reference to train the supervised ML models. VM counts were classified by intensity based on the newly established three sets of wrist cut-points and the hip reference to examine the classification accuracy. Estimates of PA were compared to the hip reference at daily level. Three thousand and six hundred epochs with matched hip- and wrist-worn accelerometer VM counts were analyzed. PA intensity cut-points for wrist-worn accelerometers varied significantly among the ML approaches. The k-means cluster analysis derived cut-points, which were: ≤ 2556 counts per minute (cpm) for SED, 2557-7064 cpm for LPA, 7065-14532 for MPA, and ≥ 14533 cpm for VPA, had the highest classification accuracy, with more than 70% of the total epochs being classified into the correct PA categories as examined by the hip reference. K-means cluster analysis derived cut-points also exhibited the most accurate estimates on SED, LPA, and VPA as the hip reference, whereas none of the three ML approaches was able to assess MPA accurately. The present study demonstrates the potential of ML approaches on establishing cut-points for wrist-worn accelerometry to assess PA in preschoolers. Future studies should replicate the findings from this study against objective observations of physical activities in preschooler in a larger sample.