Evaluating Machine Learning-Based Feature Selection Methods for Diagnosing Parkinson's Disease Under the SVM Framework

Parkinson's disease (PD) diagnosis is often challenging, leading to delayed treatment and reduced patient quality of life. This study investigates the use of machine learning, specifically Support Vector Machine (SVM), in early PD detection through dysphonia measurements via phonatory voice analysis. Two feature selection methods, L1-norm Support Vector Machine (L1-norm SVM) and Recursive Feature Elimination (RFE), are compared within the SVM framework with the data from the UCI Machine Learning Repository to improve PD classification based on dysphonia measures from sustained vowels. Results show that incorporating these techniques enhances classifier performance, with models incorporating feature selection consistently performing better. Particularly, L1-norm SVM proves more effective for PD diagnosis than RFE, demonstrating superior performance in both positive recall score and F1-macro score. Additionally, feature selection results show the significance of non-linear dysphonia measures and the importance of emphasizing both periodic and noise-related dysphonia measures in PD diagnosis. These findings suggest the potential for more accurate PD diagnosis methods in phonatory voice analysis and provide insights into enhancing PD diagnosis through machine learning and voice analysis. Further research into feature selection methods, datasets, and diagnosis algorithms is recommended to enhance early PD diagnosis performance.