Comparison of linear and non-linear models for predicting energy expenditure from raw accelerometer data

Abstract

This study had three purposes, all related to evaluating energy expenditure (EE) prediction accuracy from body-worn accelerometers: (1) compare linear regression to linear mixed models, (2) compare linear models to artificial neural network models, and (3) compare accuracy of accelerometers placed on the hip, thigh, and wrists. Forty individuals performed 13 activities in a 90 min semi-structured, laboratory-based protocol. Participants wore accelerometers on the right hip, right thigh, and both wrists and a portable metabolic analyzer (EE criterion). Four EE prediction models were developed for each accelerometer: linear regression, linear mixed, and two ANN models. EE prediction accuracy was assessed using correlations, root mean square error (RMSE), and bias and was compared across models and accelerometers using repeated-measures analysis of variance. For all accelerometer placements, there were no significant differences for correlations or RMSE between linear regression and linear mixed models (correlations: r = 0.71-0.88, RMSE: 1.11-1.61 METs; p > 0.05). For the thigh-worn accelerometer, there were no differences in correlations or RMSE between linear and ANN models (ANN-correlations: r = 0.89, RMSE: 1.07-1.08 METs. Linear models-correlations: r = 0.88, RMSE: 1.10-1.11 METs; p > 0.05). Conversely, one ANN had higher correlations and lower RMSE than both linear models for the hip (ANN-correlation: r = 0.88, RMSE: 1.12 METs. Linear models-correlations: r = 0.86, RMSE: 1.18-1.19 METs; p < 0.05), and both ANNs had higher correlations and lower RMSE than both linear models for the wrist-worn accelerometers (ANN-correlations: r = 0.82-0.84, RMSE: 1.26-1.32 METs. Linear models-correlations: r = 0.71-0.73, RMSE: 1.55-1.61 METs; p < 0.01). For studies using wrist-worn accelerometers, machine learning models offer a significant improvement in EE prediction accuracy over linear models. Conversely, linear models showed similar EE prediction accuracy to machine learning models for hip- and thigh-worn accelerometers and may be viable alternative modeling techniques for EE prediction for hip- or thigh-worn accelerometers.

Measurement of physical activity in clinical practice using accelerometers

“…This clearly limits comparability between data captured from different body positions. However, as previously mentioned, machine learning improves the accuracy of wrist measurements to a similar level as data collected at the hip .…”

Estimating energy expenditure from wrist and thigh accelerometry in free-living adults: a doubly labelled water study

“…Some have suggested that simple movement intensity approaches should be replaced by more sophisticated models that utilise a broader range of signal features 40,41 . Recent efforts to estimate energy expenditure have utilised a range of machine learning approaches, such as neural networks 42–44 and random forests 40 .…”

Re‐examination of accelerometer data processing and calibration for the assessment of physical activity intensity

“…20 Due to the complexity of arm movements, machine-learning algorithms may be required to achieve the same level of accuracy for activity intensity as with the waist and thigh location. 21 During walking, both the vertical (W EXT-VERT ) and the horizontal (W EXT-HOR ) components of W EXT increase with speed. 18,19 At the start of running, there is a large increase in W EXT-VERT , but it reaches a plateau with faster running speed, while W EXT-HOR continues to increase.…”

“…Therefore, for example, discrepancies between the wrist and waist locations can occur for free‐living measurements because their movements are being decoupled . Due to the complexity of arm movements, machine‐learning algorithms may be required to achieve the same level of accuracy for activity intensity as with the waist and thigh location …”