Predictors of Extreme Behaviors of Physical Activity Measured With Accelerometer-A Predictive Machine Learning Approach.

Abstract

Objectives: This study applied a robust predictive machine learning framework to a comprehensive set of individual background data to detect variables that are predictive of extremely low and high amounts of moderate to vigorous physical activity (MVPA). The main aim was to detect predictive variables, and as a secondary aim, the model accuracy was also assessed. Additionally, the analysis framework is presented for further use in physical activity research.

Methods: Comprehensive data on health, physical functioning, education, work, and the living environment, represented by 130 variables, from 1536 participants aged from 20 to 69 years were used for analysis. The 2.5th and 97.5th quantiles of MVPA were chosen to represent the extreme physical behaviors in the population. Random forest classifiers were trained separately to predict the low and high MVPA groups. Results were confirmed and predictive variables detected with permutation tests and the Wilcoxon signed-rank test, Bonferroni-corrected for P < .05.

Results: Nine significant predictors were found for the extremely low MVPA and 6 for the extremely high MVPA. These variables were related to the individual's mobility function, living environment, and occupation. The areas under the receiver operating characteristic curve were 0.73 and 0.59 for the participants with extremely low and high amounts of MVPA, respectively.

Conclusions: Several significant predictive variables for extremely low and high MVPA were discovered that may help characterize factors associated with PA behavior at the population level. However, it is not possible to accurately identify physically highly inactive or highly active individuals only from comprehensive background data.