Related Experiment Video
Updated: Jul 9, 2026
Visualization of Intensity Levels to Reduce the Gap Between Self-Reported and Directly Measured Physical Activity
Published on: March 7, 2019
Understanding Drivers of Physical Activity Through Explainable Machine Learning: The Role of Disability and Social
Aditya Chakraborty1, Emily J Nicklett2
1Department of Epidemiology, Biostatistics, and Environmental Health, Joint School of Public Health, Old Dominion University, Norfolk, VA, USA.
Background:
Physical activity among adults with disabilities is influenced by functional limitations, health status, and socioeconomic conditions; yet, the relative predictive importance of these factors remains insufficiently understood. This study compared multiple machine learning approaches for predicting exercise participation and used explainable artificial intelligence to identify the most influential predictors.
Methods:
Using the Centers for Disease Control and Prevention Behavioral Risk Factor Surveillance System data, we modeled self-reported exercise participation from 5 disability indicators-blindness, deafness, cognitive disability, self-care disability, and mobility disability, along with demographic, socioeconomic, behavioral, psychosocial, and self-rated health factors. Five machine learning models were compared: logistic regression, LASSO, support vector machine, random forest, and XGBoost. Performance was evaluated with AUC, accuracy, F1 score, sensitivity, specificity, and Cohen kappa. Exercise participation was also summarized across disability types by income and education. Model explainability for the best performing model was assessed using SHAP plots.
Results:
XGBoost demonstrated the strongest overall performance (AUC = 0.83, accuracy = 0.79, F1 = 0.78, sensitivity = 0.77, specificity = 0.80, κ = .54), followed by random forest (AUC = 0.80). Exercise participation showed a pronounced socioeconomic gradient within disability groups. Among respondents with cognitive disability, exercise participation declined from 75.2% in the high-income group to 40.7% in the low-income group, while among those with mobility disability, it declined from 51.7% to 29.9%. SHAP analyses identified mobility status, education attainment, income, physical health, and age as the top 5 contributors to exercise participation.
Conclusions:
Explainable machine learning may improve identification of individuals at elevated risk of inactivity and highlights the combined importance of disability type and socioeconomic context in shaping exercise participation.
