Machine Learning Based Stepping Filter Improves Estimates of Moderate to Vigorous Intensity Physical Activity from Wrist Actigraphy

In this proof-of-concept study involving older adults with varying levels of function, we sought to answer the question: can the application of a simple stepping classification algorithm reduce non-stepping sources of acceleration from wrist-worn ActiGraph data prior to the application of acceleration cut-points to quantify time spent in moderate to vigorous intensity physical activity (MVPA)? Participants completed a series of known tasks including stepping tasks, cycling tasks, and home chores repeated across two days while wearing an ActivPAL (PAL Technologies LLC, Glasgow, Scotland) activity monitor on the thigh—which provides an accurate measurement of stepping behavior—and an ActiGraph GT3X+ wrist worn device. A stepping classifier was trained for each monitor, and we examined time spent in MVPA prior to and following removal of non-stepping behaviors. Thirty participants (72.51±6.93 years; 50% female) completed the study. Across both visits, the ActiGraph reported participants engaged in 48.47±17.15 minutes of MVPA, and the ActivPAL recorded 14.00±6.44 minutes; an approximately 3.5-fold difference. After application of the classifier-based filter, the ActiGraph recorded a mean MVPA duration of 20.57±10.27 minutes of MVPA (a reduction of 60%), and the ActivPAL recorded 11.80±5.27 minutes of MVPA (a reduction of 14%); a 1.74-fold difference. These results suggest that identifying and removing non-stepping behaviors prior to scoring accelerometer data via intensity thresholds may help to reduce inflated MVPA duration attributable to non-stepping wrist movement. Future work should focus on producing strong population-specific step classification algorithms and explore their application to existing and future studies focused on PA behavior.