Detecting Parkinson’s disease with sustained phonation and speech signals using machine learning techniques

Abstract

This study investigates the processing of voice signals for detecting Parkinson’s disease. This disease is one of the neurological disorders that affect people in the world most. The approach evaluates the use of eighteen feature extraction techniques and four machine learning methods to classify data obtained from sustained phonation and speech tasks. Phonation relates to the vowel /a/ voicing task and speech to the pronunciation of a short sentence in Lithuanian language. The audio tasks were recorded using two microphone channels from acoustic cardioid (AC) and a smartphone (SP), thus allowing to evaluate the performance for different types of microphones. Five metrics were employed to analyze the classification performance: Equal Error Rate (EER) and Area Under Curve (AUC) measures from Detection Error Tradeoff (DET) and Receiver Operating Characteristic curves, Accuracy, Specificity, and Sensitivity. We compare this approach with other approaches that use the same data set. We show that the task of phonation was more efficient than speech tasks in the detection of disease. The best performance for the AC channel achieved an accuracy of 94.55%, AUC 0.87, and EER 19.01%. When using the SP channel, we have achieved an accuracy of 92.94%, AUC 0.92, and EER 14.15%.