Distinguishing among standing postures with machine learning-based classification algorithms.

The purpose of our study was to evaluate the accuracy with which classification algorithms could distinguish among standing postures based on center-of-pressure (CoP) trajectories. We performed a secondary analysis of published data from three studies: Study A) assessment of balance control on firm or foam surfaces with eyes-open or closed, Study B) quantification of postural sway in forward-backward and side-to-side directions during four standing-balance tasks that differed in difficulty, and Study C) an evaluation of the impact of two modes of transcutaneous electrical nerve stimulation on balance control in older adults. Three classification algorithms (decision tree, random forest, and k-nearest neighbor) were used to classify standing postures based on the extracted features from CoP trajectories in both the time and time-frequency domains. Such classifications enable the identification of differences and similarities in control strategy. Our results, especially those involving time-frequency features, demonstrated that distinct CoP trajectories could be identified from the extracted features in all conditions and postures in each study. Although the overall classification accuracy was similar using time-frequency features (~ 86%) for the three studies, there were substantial differences in accuracy across conditions and postures in Studies A and B but not in Study C. Nonetheless, the models were far superior to the published results with conventional metrics in distinguishing between the conditions and postures. Moreover, a Shapley Additive exPlanation analysis was able to identify the most important features that contributed to the classification performance of the models.