Voice Features Examination for Parkinson’s Disease Detection Utilizing Machine Learning Methods

Manageable symptoms can be a critical and important thing for people with Parkinson’s disease (PWP) in order to maintaining their quality of life. PWP early detection and monitoring is one of way to understand whether the medication dosage and physical therapy manage to maintain the stage of symptoms. The cheapest monitoring method can be done is based on voice signal. PD detection method based on voice signal shows promising future to be implemented into real world through an online and mobile based medical related application. However, It is necessary to select the important voice features which contribute to highest detection accuracy so that the implementation of detection through the application can be done effectively without requiring complex mathematical code. This study aim to evaluate and analize the highest accuracy and suitable voice feature related to PWP early detection and monitoring using machine learning method. Voice data recorded from study participants consist of Hughes-based stages of Parkinson’s disease (PD) patients and healthy subjects. Participants recorded their voice said ‘aaaa..’ for 5 s then the voice data calculated into 22 voice features. Those features then examine using machine learning methods such as logistic regression, random forest, KNN and deep learning CNN and classified into four classes based on Hughes standard e.g. healthy, possible, probable and definite. The experimental result showed that the most suitable features are 11 features out of 22 features which examined using random forest and CNN method contributed to highest accuracy value of 95%. This most important features then can be implemented along with CNN method into an online and mobile based application for future study.