Energy Expenditure Estimation in Children by Activity-Specific Regressions, Random Forest and Regression Trees from Raw Accelerometer Data.

The aim of the study was to compare activity-specific regressions (ASR), random forest (RFEE) and regression trees (treeEE) as methods to determine energy expenditure (EE) in children from raw accelerometer data. 41 children (age: 9.9 ± 2.2y) preformed the activities sitting, standing, walking, running, jumping, crawling, cycling and riding a scooter for 3.5 min., while 30Hz raw accelerations were collected with one tri-axial hip-accelerometer (ActigGraph GT3X) and EE was measured with a gas analyzer (Cortex MetaMax 3B). 42 different features were calculated over 1-s windows and evaluated according to their importance to predict EE. The ASR accurately predicted the EE of six activities. The ASR-biases were for sitting, standing, walking and crawling within 0.17 MET. RFEE precisely estimated the EE of cycling, riding a scooter, jumping and running with biases of -0.18, -0.21, -0.57 and -0.29 MET, respectively. The treeEE accurately predicted the EE of running and cycling (bias: -0.17 and -0.38 MET). The ASR predicted EE more accurately than RFEE or the treeEE. Using activity-specific information seems therefore to lead to more accurate results. ASR might therefore be preferred to assess EE in children with raw accelerometer data in the future. KEYWORDS: ENERGY EXPENDITURE, RAW ACCELEROMETER DATA, CLASSIFICATION, CHILDREN, DATA MINING