Classifying Parkinson's Disease Using Speech Features

Abstract

This scientific paper focuses on the detection of Parkinson's Disease in individuals using vocal/speech features. Specifically, the study records an individual's pronunciation of the short ‘a’ sound (\\ \\\ \\\ae) and applies various Signal Processing Algorithms to process the recorded data. To address the data's class imbalance, both oversampling and undersampling techniques are implemented, and their results are compared. Additionally, a feature selection algorithm is employed to eliminate noise from the dataset while retaining the relevant features necessary for the model. Logistic regression, Support Vector Machines (SVM), decision trees, and Ensemble learning methods are utilized to determine the likelihood of an individual having Parkinson's Disease. Furthermore, SHAP (SHapley Additive exPlanations) analysis is conducted to gain a deeper understanding of the impact of features on class determination for each instance. The Random Forest model achieves an accuracy of approximately 95% when trained on the oversampled dataset, surpassing the accuracy obtained from the undersampled dataset. Additionally, a graphical user interface (GUI) is successfully developed to predict the diagnosis of Parkinson's disease.