Energy expenditure prediction using raw accelerometer data in simulated free living
Montoye, 2015
Description
Sixteen artificial neural networks were developed for estimating energy expenditure from a thigh- or hip-worn ActiGraph or a GENEActiv worn on either wrist and four different sets of acceleration-based input features. Adults performed simulated activities of daily living, locomotion, and other exercises and the model was cross-validated using leave-one-out cross validation.
The following models are available. Root Mean Square Error (RMSE) is also provided.
Artificial neural network for hip ActiGraph with first feature set (1.13 METs)
Artificial neural network for hip ActiGraph with second feature set (1.32 METs)
Artificial neural network for hip ActiGraph with third feature set (1.17 METs)
Artificial neural network for hip ActiGraph with fourth feature set (1.20 METs)
Artificial neural network for thigh ActiGraph with first feature set (1.04 METs)
Artificial neural network for thigh ActiGraph with second feature set (1.08 METs)
Artificial neural network for thigh ActiGraph with third feature set (1.04 METs)
Artificial neural network for thigh ActiGraph with fourth feature set (1.04 METs)
Artificial neural network for GENEActiv on right wrist with first feature set (1.18 METs)
Artificial neural network for GENEActiv on right wrist with second feature set (1.25 METs)
Artificial neural network for GENEActiv on right wrist with third feature set (1.27 METs)
Artificial neural network for GENEActiv on right wrist with fourth feature set (1.21 METs)
Artificial neural network for GENEActiv on left wrist with first feature set (1.18 METs)
Artificial neural network for GENEActiv on left wrist with second feature set (1.26 METs)
Artificial neural network for GENEActiv on left wrist with third feature set (1.26 METs)
Artificial neural network for GENEActiv on left wrist with fourth feature set (1.15 METs)
Development/Validation
Sample: 39 (19 males) healthy adults 18-44 years of age
Setting: Laboratory
Activities: Activities of daily living, cycling, overground/self-paced walking and running, resistance exercise, stairs, stationary behaviors
Criterion: Oxycon Mobile (VO2)
Accelerometer(s): ActiGraph GT3X+ on right thigh and hip, GENEActiv on each wrist
Validation approach: Leave-one-out cross-validation
Instructions
The R code can be used to read in a csv file with the necessary features and then use the chosen model to predict energy expenditure in METs for each epoch. There is a separate file for each of the sixteen models. There is an example data sheet with the variables described in the data dictionary. Since not all monitors and/or features are used in each model, you will only need the variables for the appropriate monitor and/or features, not every variable present in the example data sheet. Further instructions are in the R code. More information about R is found here.
This code relies on R software which can be downloaded for free at https://www.r-project.org/
Attached Files
Download all files as a .zip or download individual files below.
GENEActiv right wrist feature set 1
GENEActiv right wrist feature set 2
GENEActiv right wrist feature set 3
GENEActiv right wrist feature set 4
GENEActiv left wrist feature set 1
GENEActiv left wrist feature set 2
GENEActiv left wrist feature set 3
GENEActiv left wrist feature set 4
Note: Compared to the original source, code was edited for consistency, example data and a data dictionary were added, models were unchanged
Source Information
Reference:
Montoye, A. H., Mudd, L. M., Biswas, S., & Pfeiffer, K. A. (2015). Energy expenditure prediction using raw accelerometer data in simulated free living. Medicine and Science in Sports and Exercise, 47(8), 1735-1746. https://doi.org/10.1249/mss.0000000000000597 Link to paper
Original source for code:
https://sites.google.com/site/alexmontoye/machine-learning-model-code
Corresponding author: Alexander H.K. Montoye, montoyeah@alma.edu
Contact
Kimberly Clevenger at accelerometerrepository@gmail.com