Leibniz Institute for Prevention Research and Epidemiology – BIPS
Join us on March 2nd for an ActiGraph webinar:
Wearable Data Gone Awry: Cautionary Tales from the Clinical Research Trenches.Register Now
Accelerometry-Based Prediction of Energy Expenditure in Preschoolers
- Published on June 18, 2019
Study purposes were to develop energy expenditure (EE) prediction models from raw accelerometer data and to investigate the performance of three different accelerometers on five different wear positions in preschoolers.
Fourty-one children (54% boys; 3–6.3 years) wore two Actigraph GT3X (left and right hip), three GENEActiv (right hip, left and right wrist), and one activPAL (right thigh) while completing a semi-structured protocol of 10 age-appropriate activities. Participants wore a portable indirect calorimeter to estimate EE. Utilized models to estimate EE included a linear model (LM), a mixed linear model (MLM), a random forest model (RF), and an artificial neural network model (ANN). For each accelerometer, model, and wear position, we assessed prediction accuracy via leave-one-out cross-validation and calculated the root-mean-squared-error (RMSE).
Mean RMSE ranged from 2.56–2.76 kJ/min for the RF, 2.72–3.08 kJ/min for the ANN, 2.83–2.94 kJ/min for the LM, and 2.81–2.92 kJ/min for the MLM. The GENEActive obtained mean RMSE of 2.56 kJ/min (left and right wrist) and 2.73 kJ/min (right hip). Predicting EE using the GT3X on the left and right hip obtained mean RMSE of 2.60 and 2.74 kJ/min. The activPAL obtained a mean RMSE of 2.76 kJ/min.
These results demonstrate good prediction accuracy for recent accelerometers on different wear positions in preschoolers. The RF and ANN were equally accurate in EE prediction compared with (mixed) linear models. The RF seems to be a viable alternative to linear and ANN models for EE prediction in young children in a semi-structured setting.