Automated Classification of Postural Control for Individuals With Parkinson’s Disease Using a Machine Learning Approach: A Preliminary Study

Abstract

The purposes of the study were (1) to compare postural sway between participants with Parkinson’s disease (PD) and healthy controls and (2) to develop and validate an automated classification of PD postural control patterns using a machine learning approach. A total of 9 participants in the early stage of PD and 12 healthy controls were recruited. Participants were instructed to stand on a force plate and maintain stillness for 2 minutes with eyes open and eyes closed. The center of pressure data were collected at 50 Hz. Linear displacements, standard deviations, total distances, sway areas, and multiscale entropy of center of pressure were calculated and compared using mixed-model analysis of variance. Five supervised machine learning algorithms (ie, logistic regression, K-nearest neighbors, Naïve Bayes, decision trees, and random forest) were used to classify PD postural control patterns. Participants with PD exhibited greater center of pressure sway and variability compared with controls. The K-nearest neighbor method exhibited the best prediction performance with an accuracy rate of up to 0.86. In conclusion, participants with PD exhibited impaired postural stability and their postural sway features could be identified by machine learning algorithms.

Distinguishing features of postural instability in Parkinson’s disease identified through wireless insole plantar pressure monitoring during daily living activities

“…There was also a large distribution of this feature seen in individuals with PD as compared to controls. These findings align with previous studies that found machine learning algorithms could classify PD subjects from control subjects based on postural sway features 25,32 . Specifically, postural sway was greater and more variable in PD subjects compared to controls 25 , and that rms of the COP was a useful feature in the differentiation 32 .…”

“…Previous studies that have leveraged machine learning models to classify PD and control groups have focused primarily on one task, including quiet stance (86% accuracy) 25 or gait (95% accuracy) 26 . The improved performance achieved by static+active tasks models in this study likely stems from choosing tasks that sample non-redundant aspects of postural control.…”

Model of the Performance Based on Artificial Intelligence–Fuzzy Logic Description of Physical Activity

“…The only study of postural control using artificial intelligence that we are aware of is the development and validation of the automatic identification of postural control patterns in children with autism spectrum disorders using a machine learning approach (naive Bayes method) [ 30 ]. These are similar to those used in the Parkinson’s disease group, where among the five supervised machine learning algorithms (logistic regression, K-nearest neighbors, naive Bayes, decision trees and random forest), the kNN method showed the best results for predicting impaired postural stability [ 31 ]. However, this did not involve athletes, nor did it use fuzzy logic, so it is difficult to compare such distant studies other than to demonstrate the usefulness of machine learning for analyzing postural control in the above-mentioned group of children [ 30 ].…”

Distinguishing features of Parkinson’s disease fallers based on wireless insole plantar pressure monitoring

“…Previous studies that have leveraged machine-learning models to classify PD and control groups have focused primarily on one task, including quiet stance (86% accuracy) 26 or gait (95% accuracy) 27 . The improved performance achieved by static+active tasks models in this study likely stems from choosing tasks that sample non-redundant aspects of postural control.…”