Automatic Classification of Parkinson's Disease Using Best Parameters of Forward and Backward Walking

Abstract

This study aims to investigate the discriminative gait features of forward and backward walking to provide a combination of the most relevant parameters. These parameters would potentially help the clinicians to follow quantitative methods in diagnosing Parkinson's disease. In this paper, the statistically significant gait features were narrowed down from 46 to 30, 20, 10, and 5, using the minimal-redundancy-maximal-relevance feature selection method. The selected features were then fed to Random Forest and Support Vector Machine classifiers to evaluate the ability of features in discriminating Parkinson's disease and control groups. According to the results, we selected to use Random Forest classifier in our algorithm. Applying our algorithm on a database comprising 62 Parkinson's disease patients and 11 control participants, we achieved the average accuracy of 93.9 and 88 in 10 iterations of Random Forest and Support Vector Machine, respectively. Using the minimal-redundancy-maximal-relevance feature selection and mean decrease in accuracy and Gini index of the Random Forest classifier, we found the critical role of backward walking parameters such as the average of stance time, step length, and swing time in classification results.