Improving energy expenditure estimates from wearable devices: A machine learning approach

A means of quantifying continuous, free-living energy expenditure (EE) would advance the study of bioenergetics. The aim of this study was to apply a non-linear, machine learning algorithm (random forest) to predict minute level EE for a range of activities using acceleration, physiological signals (e.g., heart rate, body temperature, galvanic skin response), and participant characteristics (e.g., sex, age, height, weight, body composition) collected from wearable devices (Fitbit charge 2, Polar H7, SenseWear Armband Mini and Actigraph GT3-x) as potential inputs. By utilising a leave-one-out cross-validation approach in 59 subjects, we investigated the predictive accuracy in sedentary, ambulatory, household, and cycling activities compared to indirect calorimetry (Vyntus CPX). Over all activities, correlations of at least r = 0.85 were achieved by the models. Root mean squared error ranged from 1 to 1.37 METs and all overall models were statistically equivalent to the criterion measure. Significantly lower error was observed for Actigraph and Sensewear models, when compared to the manufacturer provided estimates of the Sensewear Armband (p < 0.05). A high degree of accuracy in EE estimation was achieved by applying non-linear models to wearable devices which may offer a means to capture the energy cost of free-living activities.