Diagnosis of Parkinson's Disease using Principal Component Analysis and Machine Learning algorithms with Vocal Features
Dhulipalla Venkata Rao, Y Sucharitha, D Venkatesh, K Mahamthy, SK MD Yasin
Abstract
This paper aims to propose a diagnosis model to detect Parkinson's Disease using Principal Component Analysis (PCA) and Random Forest using the extracted speech features from the voice recordings of both PD and HC. Firstly, we have conducted a comparative analysis of Machine Learning algorithms like Support Vector Machine (SVM), K Nearest Neighbor (KNN), Random Forest (RF) without feature reduction. Principal Component Analysis is an unsupervised dimensionality reduction technique to remove redundant data and speed up the training, testing process. Secondly, we have used PCA to reduce 26 voice features into two principal components that represents more information and highly correlated to output. Then, the optimized data is fed through the Machine Learning algorithms such as SVM, KNN, Random Forest to classify PD and HC. The results show that Random Forest has outperformed the remaining algorithms with an accuracy of 95%, F1 score of 0.969 and AUC score of 0.977. The y-axis value of ROC curve approaches to 1 for Random Forest Algorithm. The proposed model improves the F1 score, AUC and practically classify the PD and HC patients.