Feature Selection Techniques Applied to Voice-based Prediction of Parkinson's Disease

Parkinson's disease is a common neurodegenerative disorder where the typical clinical features are manifested as muscle stiffness, bradykinesia, and resting tremors. Voice or speech features appear prominent in the early phase of the disease and hence can be utilized as a non-invasive and a very simple way of assessing the presence of Parkinson's disease. The goal of the study is to predict the occurrence of Parkinson's disease in its early phase from voice or speech signals through different machine learning algorithms namely: k-Nearest Neighbors, random forest, support vector machine, logistic regression, and Adaptive Boosting. This research also aims to compare the contributions of different feature selection procedures and its effect on machine-learning algorithms for PD prediction. The overall topmost models were obtained by support vector machine with feature selection by Pearson r (utilizing 12 acoustic features) and logistic regression with feature selection by F-score (utilizing 18 acoustic features). These models achieved the highest accuracy rate of 87.50%, 85.71% recall, 92.31% precision and 88.95% F1-score. The results suggested that acoustic signals from voice recordings can reliably discriminate Parkinson's disease even in its initial stage. The prompt detection using machine learning and its integration in the clinical workflow of the health professionals could be harnessed as a non-invasive means of detecting Parkinson's disease. This would enable swift administration of therapeutic intervention to mitigate disease progression leading to a prolonged period of optimal functioning capacity. All of these consequently ameliorate the overall well-being of those afflicted with Parkinson's disease.