Registration Is Now Open!
Virtual ActiGraph Digital Data Symposium 2021November 4, 2021 | 10:30 AM - 1:00 PM CT | Learn more
Multi-group and Longitudinal Invariance of Artificial Neural Networks (ANNs) to Predict Physical Activity Type and Physical Activity Energy Expenditure in Children and Adolescents
- Added on June 15, 2012
Purpose To evaluate the multi-group and longitudinal invariance of ANNs to predict physical activity (PA) type and PA energy expenditure (METs) in youth.
Methods The study was conducted at two sites. Site 1 (S1) served as the training and testing sample and comprised 100 youth mean age 11.0 ± 2.7 y. Site 2 (S2) served as the cross-validation sample and comprised 106 youth mean age 10.8 ± 2.6 y. Participants completed 12 activity trials categorized into 5 PA types: sedentary, walking, running, light household activities or games, and moderate-to-vigorous games or sports. During each trial, participants wore an ActiGraph GT1M and VO2 was measured using the Oxycon Mobile portable metabolic system. Participants at S1 repeated the protocol after 24-months. ANNs to predict PA type and METs were trained and tested using features in the processed count data: 10th, 25th, 50th, 75th, and 90th percentiles and the lag one autocorrelation. To test multi-group invariance, the best performing networks from S1 were tested using data collected at S2. To evaluate longitudinal invariance, the best performing networks from S1 were tested in the same participants 24 months later.
Results Classification accuracy was 90.4% and 85.6% for S1 and S2, respectively. RMSE was 0.9 METs and 1.0 METs for S1 and S2, respectively. At 24-months follow-up, classification accuracy and RMSE for S1 participants was 80.5% and 1.1 METs, respectively.
Conclusions Previously trained and tested ANNs to predict PA type and intensity from accelerometer output perform well when applied to other children and the same children 2 years later.