Quantifying the uncertainty of human activity recognition using a Bayesian machine learning method: a prediction study.

PURPOSE: Machine learning methods accurately predict physical activity outcomes using accelerometer data generated by wearable devices. Thus, they allow investigation of the impact of the built environment on population physical activity. Although traditional machine learning methods do not provide prediction uncertainty, a new method, Bayesian Additive Regression Trees (BART), can quantify such uncertainty as a posterior predictive distribution. Our objective was to evaluate the performance of BART in regard to predicting physical activity status.

METHODS: We applied multinomial BART and a benchmark method, random forest, to accelerometer data at 25,424 time points generated by wearable devices worn by 37 participants. We evaluated the prediction accuracy, F1 scores, and confusion matrices using leave-one-person-out cross-validation.

RESULTS: BART and random forest demonstrated comparable prediction performances.

CONCLUSION: BART is a relatively novel machine learning method that can advance the incorporation of the predicted physical activity status into built environment research. Future research should evaluate the association between the environment and physical activity as predicted by BART.