Validation and Comparison of Accelerometers Worn on the Hip, Thigh, and Wrists for Measuring Physical Activity and Sedentary Behavior

Abstract

Background: Recent evidence suggests that physical activity (PA) and sedentary behavior (SB) exert independent effects on health. Therefore, measurement methods that can accurately assess both constructs are needed.

Objective: To compare the accuracy of accelerometers placed on the hip, thigh, and wrists, coupled with machine learning models, for measurement of PA intensity category (SB, light-intensity PA [LPA], and moderate- to vigorous-intensity PA [MVPA]) and breaks in SB.

Methods: Forty young adults (21 female; age 22.0 ± 4.2 years) participated in a 90-minute semi-structured protocol, performing 13 activities (three sedentary, 10 non-sedentary) for 3-10 minutes each. Participants chose activity order, duration, and intensity. Direct observation (DO) was used as a criterion measure of PA intensity category, and transitions from SB to a non-sedentary activity were breaks in SB. Participants wore four accelerometers (right hip, right thigh, and both wrists), and a machine learning model was created for each accelerometer to predict PA intensity category. Sensitivity and specificity for PA intensity category classification were calculated and compared across accelerometers using repeated measures analysis of variance, and the number of breaks in SB was compared using repeated measures analysis of variance.

Results: Sensitivity and specificity values for the thigh-worn accelerometer were higher than for wrist- or hip-worn accelerometers, > 99% for all PA intensity categories. Sensitivity and specificity for the hip-worn accelerometer were 87-95% and 93-97%. The left wrist-worn accelerometer had sensitivities and specificities of > 97% for SB and LPA and 91-95% for MVPA, whereas the right wrist-worn accelerometer had sensitivities and specificities of 93-99% for SB and LPA but 67-84% for MVPA. The thigh-worn accelerometer had high accuracy for breaks in SB; all other accelerometers overestimated breaks in SB.

Conclusion: Coupled with machine learning modeling, the thigh-worn accelerometer should be considered when objectively assessing PA and SB.

Keywords: activity monitor; activity tracker; artificial neural network; energy expenditure; machine learning; pattern recognition.

Conflict of interest statement

Conflict of Interest: The authors attest that they have no conflicts of interest to report.

Figures

Figure 1.. ANN for predicting PA intensity category.

*The number of input features was 15 (5 features*3 measurement axes). Abbreviations: 10th % ile: 10th percentile of acceleration signal. 25th % ile: 25th percentile of acceleration signal. 50th % ile: 50th percentile of acceleration signal. 75th % ile: 75th percentile of acceleration signal. 90th % ile: 90th percentile of acceleration signal. S: summation functions of the input layer in the hidden units. U: activation function for the hidden layer. W1: the weight vectors for each of the inputs. W2: the weight vectors for each of the summations.

Figure 2.. Confusion matrices for prediction of SB, LPA, and MVPA.

For a-d, rows are actual PA intensities and columns are predicted PA intensities. Grey boxes represent number of instances where the PA intensity category was correctly predicted.

Figure 3.. Predicted vs. measured time in each PA intensity category.

Error bars represent standard deviation. * Indicates significant differences from the criterion measure (direct observation).

Figure 4.. Predicted vs. measured breaks in SB.

Error bars represent standard deviation. * Indicates significant differences from the criterion measure (direct observation).