Parkinson’s Disease Severity Estimation on Hungarian Speech Using Various Speech Tasks

Abstract

In the present paper, classification and severity level estimation are carried out based on speech applying various speech tasks in order to help developing a method to diagnose the disease in an early stage. A database is introduced, which contains speech samples uttered by Hungarian Parkinson patients and healthy control population. Classification and regression models are built using various machine-learning methods for all speech tasks separately. Additional to the separate decisions for all speech tasks, a joint decision was made for each speaker. The final prediction was obtained by fusing the separate estimations for each speaker. Test trials were run in order to investigate, if age is a relevant feature for the machine learning tasks. The best results were obtained using support vector machine with 89.3% accuracy for binary classification. With regression, the method achieved 0.787 Spearman correlation for Parkinson severity level estimation measured on Hoehn-Yahr scale. Performance measures slightly varied according to gender differences. The investigated features seem to be useful and relevant for the automatic diagnosis of Parkinson’s disease based on the classification and regression performances.