Machine Learning Based Automatic Prediction of Parkinson’s Disease Using Speech Features

Abstract

Parkinson’s disease is a severe neurodegenerative disease where primarily the motor system of the human body gets affected. It is currently one of the leading causes of disability around the world. Although currently there does not exist any cure for the disease, early detection of the disease can lead to better treatment and better disease management, which can improve the quality of life of the patients drastically. Researchers have shown that speech data can be very helpful in the early diagnosis of the disease. Clinical decision support systems built using speech features from previous patients could potentially be of very high utility for early diagnosis of the disease. In this work, a machine learning based automatic prediction framework is presented, which can be used to build such effective decision support systems. Various evaluation metrics like prediction accuracy, sensitivity, specificity and Area Under the Curve of the Receiver Operating Characteristics have been considered for evaluating the applicability of various machine learning algorithms. It has been observed that ensemble methods like Random Forests and Gradient Boosting classifiers can be used to effectively perform predictive modelling, achieving average prediction accuracy of 86.5%. Oversampling strategy is also used in order to increase the size of the training data, so that deep neural network based approaches can be studied. Significant improvements have been observed post oversampling, achieving average prediction accuracy up to 91.5, which suggests the potential for applicability of the approach as a decision support system in real diagnostic scenarios.