Deep learning to predict energy expenditure and activity intensity in free living conditions using wrist-specific accelerometry

Wrist-worn accelerometers are more comfortable and yield greater compliance than hip-worn devices, making them attractive for free-living activity assessments. However, intricate wrist movements may require more complex predictive models than those applied to hip-worn devices. This study developed a novel deep learning method that predicts energy expenditure and physical activity intensity of adults using wrist-specific accelerometry. Triaxial accelerometers were worn by 119 participants on their wrist and hip for two weeks during waking hours. A deep learning model was developed from week 1 data of 60 participants and tested using week 2 data for: (i) the remaining 59 participants (Group UT), and (ii) participants used for training (Group TR). Estimates of physical activity were compared to a reference hip-specific method. Moderate-to-vigorous physical activity predicted by the wrist-model was not different to the reference method for participants in Group UT (5.9±3.1vs. 6.3±3.3 hour/week) and Group TR (6.9±3.7 vs. 7.2±4.2 hour/week). At 60-s epoch level, energy expenditure predicted by the wrist-model on Group UT was strongly correlated with the reference method (r=0.86, 95%CI: 0.84–0.87) and closely predicted activity intensity (83.7%, 95%CI: 80.9–86.5%). The deep learning method has application for wrist-worn accelerometry in free-living adults.