A Model For Predicting Energy Consumption At Different Step Speeds In Humans Based On Wrist-Worn Accelerometers

Objective: Monitoring human motion based on a multi-axis accelerometer has a wide range of applications.To solve the problem of the low accuracy of the accelerometer in predicting energy consumption in human motion,based on the motion data collected by a human wristwearing accelerometer,the prediction models of energy consumption at different walking speeds are established through a variety of algorithms.the accuracy of the model is evaluated and a better energy consumption prediction model is selected to provide a theoretical reference for selecting a reasonable algorithm to predict energy consumption in human motion.Methods: 100 students aged 18-30 years old were selected as subjects,and the subjects were randomly divided into a model group(n = 70)and a verification group(n = 30).The subjects wore a triaxial accelerometer,COSMED Quart PFT ergo cardiopulmonary function tester(Quart PFT),and heart rate band according to a sitting posture,walking(2 km/h,3 km/h,4km/h,5 km/h,6 km/h)and running(7 km/h,8 km/h,9 km/h).Acceleration data were collected by triaxial accelerometer and energy consumption data were collected by Quart PFT.The group regression equation is established with the processed acceleration data Mean as the independent variable and metabolic equivalent(metabolic equivalent,METs)as the dependent variable.With mean,sd,max,min and five quantile(10th、25th、50th、75th、90th)of 60 s window acceleration data as input layer variables and METs as output layer variables,artificial neural network(ANN)model and two-stage model were established.The accuracy of the energy consumption prediction of the model is verified by verification group data.Results: The linear equation(METs = 8.33 Mean + 3.36),the logarithmic equation(METs= 2.56 × ln(Mean)+ 10.04),the cubic equation(METs = 29.65 Mean3-52.67 Mean2 + 33.46 Mean + 1.22),the ANN model and ANN walking-and-running two-stage model were established.The results show that there is no significant difference between the METs predicted by the five models and the measured METs in the overall energy consumption prediction accuracy of the model.The correlation between the METs predicted by the ANN walking-and-running two-stage model and the measured METs is the highest(r = 0.913),while the Root Mean Square Error(RMSE)and Bias of the ANN walking-and-running twostage model are the smallest(RMSE = 0.76 METs,Bias = 0.02 METs).In terms of consistency measurement,the ANN walking-and-running two-stage model has the least points outside the consistency interval and the highest degree of consistency.In the comparison of the prediction accuracy of energy consumption in walking and running,the RMSE and Bias of the ANN walking-and-running two-stage model were the lowest(RMSE= 0.66 METs,Bias = 0.03 METs),and the lowest RMSE was ANN walking-and-running two-stage model(RMSE = 0.90 METs).The linear equation,logarithmic equation,ANN model,and ANN walking-running two-stage model all had lower Bias(Bias < 0.01METs).In the comparison of the accuracy of energy consumption prediction under a single walking speed,the points in the ANN walking-and-running two-stage model are closer to the reference line,and the RMSE and Bias at different walking speeds are lower than the other four models as a whole.Conclusion: The ANN walking-and-running two-stage model based on wrist accelerometer has high accuracy in predicting energy consumption,which is superior to the linear equation,logarithmic equation,cubic equation and ANN model established in this study.