Ensemble Learning Classifier with Optimal Feature Selection for Parkinson's Disease

Parkinson's Disease (PD) is a most common neurological disorder leading to gait impairments. Gait being a potential biomarker evolving with time and its integration with machine learning techniques can serve as a detection method to assist medical professionals. The current study uses database of 29 PD and 25 Healthy Controls (HC) during dynamic walking on leveled ground. The primary focus of the study is to differentiate PD and healthy subjects by extracting clinically relevant features from Vertical Ground Reaction Force (VGRF) database. Recursive Feature Elimination with 10-fold Cross-Validation (RFECV) for Random Forest (RF) is used to select best possible combination of features. The prediction models built using ensemble learning classifiers reported maximum accuracy of 92.9% ± 0.73 using bagging techniques and 91.38% ± 0.79 with boosting techniques. Thus, this study identified the best suited subset of features that can enhance PD diagnosis in clinical decision making.