Statistical Learning Methods to Predict Activity Intensity from Body-Worn Accelerometers

Abstract

Physical activity, especially when performed at moderate or vigorous intensity, has short- and long-term health benefits, but measurement of free-living physical activity is challenging. Accelerometers are popular tools to assess physical activity, although accuracy of conventional accelerometer analysis methods is suboptimal. This study developed and tested statistical learning models for assessing activity intensity from body-worn accelerometers. Twenty-eight adults performed 10-21 activities of daily living in two visits while wearing four accelerometers (right hip, right ankle, both wrists). Accelerometer placement is of crucial practical concern and this paper addresses this issue. Boosting, bagging, random forest and decision tree models were created for each accelerometer and for two-, three-, and four-accelerometer combinations to predict activity intensity. Research staff observations of activity intensity served as the criterion. Point estimates of error for the ankle accelerometer were 2.2-4.7 percentage points lower than other single-accelerometer placements, and the left wrist-ankle combination had errors 0.8-5.8 percentage points lower than other two-accelerometer combinations. Decision trees had poorer accuracy than the other models. Using an accelerometer worn on the lower limb, by itself or in combination with an upper-limb accelerometer, appears to offer optimal accuracy for activity intensity measurement.

Modeling of ordinal longitudinal accelerometer data

“…One method that has been implemented is to summarize raw acceleration data into every-30second epochs (Montoye et al, 2017b,a;Lazar et al, 2020). These summaries make up new explanatory variables and consist of summaries for each axis of acceleration and summaries of pairwise associations between these axes.…”

“…These use a number of different types of statistical and prediction models as well as analyze different placements, different ways of measuring level of activity and specific uses of accelerometers in predicting physical activity levels. For example, ensemble methods based on decision trees such as bagging, boosting and random forest have been utilized in Montoye et al (2018), Ellis et al (2014), Lazar et al (2020) and Pavey et al (2017), among others. Parametric and linear models to analyze accelerometer data, including interpretation of coefficients and significance of variables, have been explored in Montoye et al (2017a), Lazar et al (2020) and Robert et al (2009).…”

“…For example, ensemble methods based on decision trees such as bagging, boosting and random forest have been utilized in Montoye et al (2018), Ellis et al (2014), Lazar et al (2020) and Pavey et al (2017), among others. Parametric and linear models to analyze accelerometer data, including interpretation of coefficients and significance of variables, have been explored in Montoye et al (2017a), Lazar et al (2020) and Robert et al (2009). Neural networks and deep learning have also been employed in De Vries et al (2011), Nawaratne et al (2020) and Montoye et al (2017a).…”