Predicting energy expenditure from a wrist-worn ActiGraph GT3X+ accelerometer

Purpose: To examine the validity of a wrist-worn accelerometer for predicting energy expenditure in adults.

Methods: Thirty-eight adults (Mage: 35.7 ± 12.2; 54% male) participated in up to 6 prescribed activities of daily living (e.g. washing dishes, sweeping), and walking (slow and brisk), while wearing an ActiGraph GT3X+ accelerometer on the non-dominant wrist. Concurrent assessment of oxygen consumption was performed using a Cosmed K4b indirect calorimeter. Each activity was performed for 6 minutes and minutes 3-5 were used in the analysis to represent steady-state activity. METS were predicted by the three movement planes and vector magnitude (VM) in counts per minute using linear mixed-effects models. A 10-fold cross validation technique was used to calculate the error rate. Activity intensity classification in the predicted model was compared to classification using Freedson-1998, Freedson-VM and Troiano-2008 hip-based cut points on wrist-worn data.

Results: The final prediction model used the vertical axis (A1) and anterior-posterior axis (A3) (rMSE = 1.63, Bias = -0.31) defining the equation as: METS = e^((1.2514+ 0.000050A1- 0.000031A3)). The predicted model has a sensitivity of 59.8% and a specificity of 59.1% for activity intensity classification and outperformed hip-cut-point-based activity intensity classification.

Conclusions: A wrist-based accelerometer has a low sensitivity and specificity for activity intensity classification. However, when hip-based cut-points are used to classify activity intensity from a wrist-worn device, validity is even poorer. Due to the complexity of arm movement during daily living activities, it may be beneficial to use machine learning to improve classification accuracy.