Use of wrist-worn accelerometer data to classify physical activity levels has not been validated in pre-school aged children.The goal was to use machine learning (ML) methods to determine cut-points and assess their accuracy compared to validated hip-worn thresholds (Butte et al, 2014).


Wrist and hip-worn accelerometer data were collected from 34 male (n=14) and female (n=20) children, aged 3 to 4 years, enrolled in a pilot study for a randomized controlled trial.Hip-worn data were classified into sedentary (SED), light (LPA) and moderate-to-vigorous (MVPA) physical activity based on previously validated cut-points.ML methods, including ordinal logistic regression (OLR), receiver operating characteristic (ROC), and K-means cluster analysis (K-means), were used to establish SED, LPA and MVPA cut-points for wrist-worn data.Results were compared to hip-based categories on percent correct classification and difference in mean daily minutes in each activity level.


Correct classification was 72.4% for OLR, 63.6% for ROC, and 65.6% for K-means. Hip-worn data yielded daily mean minutes of 146.0 for SED, 82.2 for LPA, and 32.6 for MVPA.Daily mean minutes of SED, LPA and MVPA for OLR method were 205.0 (p<0.001), 60.0 (p<0.001), and 7.6 (p<0.001); for ROC method were 155.5 (p=1.0), 47.0 (p<0.001), and 71.0 (p<0.001); for K-means method were 136.0 (p=1.0), 88.3 (p=1.0), and 48.9 (p=0.04), respectively.Mean absolute differences were 35.4 for OLR, 42.0 for ROC, and 27.6 for K-means.


This study demonstrates potential for ML techniques in establishing wrist-worn accelerometer cut-points to classify physical activity in pre-school aged children.