Modeling clinically validated physical activity using commodity hardware

Fitbit devices are one of the most popular wearable activity monitors in the consumer market. They are considerably cheaper than many of their clinical grade counterparts. However, they utilize proprietary algorithms for estimation of physical activity (PA). This study aims to model the measures of PA as reported by the ActiGraph GT3X using Fitbit measures of steps, METs, and intensity level. Such a model relating the Fitbit to what would have been reported by the ActiGraph could enable researchers to use the Fitbit instead of the ActiGraph in some applications, thus reducing cost or increasing the number of subjects involved in a study. This paper describes a study in which a model of the Freedson VM3 physical activity classification was constructed that uses measures from the Fitbit device instead of the typically provided ActiGraph vector magnitude. The data from 19 subjects, who concurrently wore both the ActiGraph and Fitbit devices for an average of 1.8 weeks, was used to generate the minute level based model. Several modeling methods were tested; a naïve Bayes classifier was chosen based on the lowest achieved error rate. That model reduces overall Fitbit to Acti-Graph errors from 19.97% to 16.32%, a notable improvement. More importantly, it reduces the errors in moderate to vigorous physical activity levels by 40%, eliminating a statistically significant difference between MVPA estimates provided by the Freedson VM3 and Fitbit Intensity scores. This justifies use of the Fitbit device in place of an ActiGraph device in some large scale studies, especially those where MVPA estimates are of importance.