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

Phase Designation

(What's this?)

This model is in Phase 0 and 2.

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.

Code

Example data

Data dictionary

ActiGraph hip feature set 1

ActiGraph hip feature set 2

ActiGraph hip feature set 3

ActiGraph hip feature set 4

ActiGraph thigh feature set 1

ActiGraph thigh feature set 2

ActiGraph thigh feature set 3

ActiGraph thigh feature set 4

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