Predicting children’s energy expenditure during physical activity using deep learning and wearable sensor data

Abstract

This study examined a series of machine learning models, evaluating their effectiveness in assessing children’s energy expenditure, in terms of the metabolic equivalents (MET) of physical activity (PA), from triaxial accelerometery. The study also determined the impact of the sensor placement (waist, ankle or wrist) on the machine learning model’s predictive performance. Twenty-eight healthy Caucasian children aged 8–11years (13 girls, 15 boys) undertook a series of activities reflective of different levels of PA (lying supine, seated and playing with Lego, slow walking, medium walking, and a medium paced run, instep passing a football, overarm throwing and catching and stationary cycling). Energy expenditure and physical activity were assessed during all activities using accelerometers (GENEActiv monitor) worn on four locations (i.e. non-dominant wrist, dominant wrist, dominant waist, dominant ankle) and breath-by-breath calorimetry data. MET values ranged from 1.2 ± 0.2 for seated playing with Lego to 4.1 ± 0.8 for running at 6.5 kmph⁻¹. Machine learning models were used to determine the MET values from the accelerometer data and to determine which placement location performed more effectively in predicting the PA data. The study identified that novel machine learning models can be used to accurately predict METs, with 90% accuracy. The models showed a preference towards the dominant wrist or ankle as the movement in those positions were more consistent during PA. It was evident that machine learning models using these locations can be effectively used to accurately predict METs for PA in children.

Keywords:

• Indirect calorimetry
• energy expenditure
• machine learning
• accelerometer
• waist
• wrist
• ankle
• GENEActiv

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplemental data

Supplemental data for this article can be accessed at https://doi.org/10.1080/17461391.2020.1789749.