Parkinson's Disease Stage Classification with Gait Analysis using Machine Learning Techniques and SMOTE-based Approach for Class Imbalance Problem

High variability in symptom severity and progression rate roots the need for a diverse training dataset, to build an efficient Parkinson's Disease (PD) severity prediction model. The Physionet database comprises gait signals of PD subjects belonging to various H&Y score-based severity levels but forms an imbalanced dataset. A dataset is said to be imbalanced if the representation of the classification categories within a dataset is not equal. The severity of misclassifying abnormal cases as normal is high and thus is a matter of concern. This paper shows how a technique called Synthetic Minority Oversampling Technique (SMOTE) deals with the class imbalance problem in PD stage-wise classification by improving minority class recognition. The method is validated by quantifying the dissimilarity among samples generated showing the non-existence of overlapping or replication. Spatiotemporal gait parameters along with their regularity and symmetry features are the attributes considered. Classifiers are trained with balanced & imbalanced datasets and their predictive accuracy attributes are compared. Results show an improvement in determining the minority class by the model trained with the balanced dataset, thus improving the generalizability of the model.