COMPARATIVE STUDY OF DIFFERENT MACHINE LEARNING MODELS FOR PARKINSON’S DISEASE DETECTION

Abstract

Parkinson’s disease (PD) is a debilitating neurodegenerative disorder with pronounced effects on motor function and daily life. Early symptoms, which are typically subtle and include stiffness of the muscles, tremor, and disturbances in balance, make early diagno sis difficult. The standard assessment tools, such as blood tests and imaging scans, offer limited value in PD diagnosis from the onset. Impairments of the voice are an early indi cator that has potential for the prediction of PD. This research utilizes biomedical voice recordings using the University of California, Irvine (UCI) dataset to construct predic tion models for PD diagnosis. Various machine learning methods being examined include Decision-Tree Model, XGBoost, Naive Bayes, Random-Forest Model, Support Vector Ma chine (SVM) Model, Logistic Regression, and K-Nearest Neighbors (KNN). The models have been trained and tested to assess how they perform. The performance of such models has been properly assessed based on accuracy, efficiency, and processing speed. A compar ative study of the best-performing models is discussed, highlighting the capability of all the models to attain good accuracy in early-stage PD detection. In addition, the present study measures the feasibility of using lightweight models for use in mobile applications that offer accessibility in actual healthcare environments. The reliability and accuracy of PD classification prediction, this study also integrates an exploration of several boost ing models, which are renowned for their efficiency in optimizing poor learners through refined iteration. Sophisticated boosting algorithms like AdaBoost, Gradient Boosting, LightGBM, CatBoost, and Stochastic Gradient Boosting were compared to conventional models simultaneously. These methods proved to be superior in terms of precision, recall, and stability, with Stochastic Gradient Boosting being the highest in overall accuracy. Their ability to work with high-dimensional and unbalanced data renders them very use ful for biomedical applications. This blending of boosting models not only enhances the predictive strength of PD detection systems but also opens the door for scalable, real-time diagnostics that can be used in clinic or remote monitoring settings.