Distinguishing Passive and Active Standing Behaviors From Accelerometry

• Published in: Journal for the measurement of physical behaviour 2024-01, Vol.7 (1)
• Main Authors: Kowalsky, Robert J., van Werkhoven, Herman, Meucci, Marco, Quinn, Tyler D., Stoner, Lee, Hearon, Christopher M., Barone Gibbs, Bethany
• Format: Article
• Language: English
• ISSN: 2575-6605; 2575-6613
• DOI: 10.1123/jmpb.2024-0004

Purpose : To investigate whether active standing can be identified separately from passive standing via accelerometry data and to develop and test the accuracy of a machine-learning model to classify active and passive standing. Methods : Ten participants wore a thigh-mounted activPAL monitor and stood for three 5-min periods in the following order: (a) PASSIVE: standing with no movement; (b) ACTIVE: five structured weight-shifting micromovements in the medial–lateral, superior–inferior, and anterior–poster planes while standing; and (c) FREE: participant’s choice of active standing. Averages of absolute resultant acceleration values in 15-s epochs were compared via analysis of variance (Bonferroni adjustment for pairwise comparisons) to confirm the dichotomization ability of the standing behaviors. Absolute resultant acceleration values and SD s in 2- and 5-s epochs were used to develop a machine-learning model using leave-one-subject-out cross validation. The final accuracy of the model was assessed using the area under the curve from a receiver operating characteristic curve. Results : Comparison of resultant accelerations across the three conditions (PASSIVE, ACTIVE, and FREE) resulted in a significant omnibus difference, F (2, 19) = [116], p < .001, η 2 = .86, and in all pairwise post hoc comparisons (all p < .001). The machine-learning model using 5-s epochs resulted in 94% accuracy for the classification of PASSIVE versus ACTIVE standing. Model application to the FREE data resulted in an absolute average difference of 4.8% versus direct observation and an area under the curve value of 0.71. Conclusions : Active standing in three planes of movement can be identified from thigh-worn accelerometry via a machine-learning model, yet model refinement is warranted.