Improved Classification with Simultaneous Feature Selection of Speech Signals for Parkinson's Patients

Abstract

Parkinson's disease (PD) is a brain disorder occurs due to a deficiency of dopamine hormone that regulates activities of the human body. Generally, the disease can be diagnosed by clinicians through clinical observation where they categorized PD patients on a PD assessment scale to understand disease severity in order to define a therapy/treatment plan. The clinicians have a view that this approach is not suitable for diagnosis at an early stage of the disease. Recent research outcome has shown that PD patients exhibit vocal impairment at the early stage of the disease, and this is now becoming a benchmark for early PD detection. Often researchers employ state-of-the-art speech analysis techniques that exploit various extracted features to meet the objective. An optimal set of features that best explains the problem often requires careful attention to the selection of extracted features in use. As a general practice, data analysts have a view that it is better to collect as many features as possible related to the problem but at the same time, it is also believed that the presence of some noisy features can also compromise classification ability. Our main objective in this work is to select/identify the optimal set of features to utilize for the machine learning classification models with an objective to have an improved early PD detection in patients. The selection of optimal features set will not only help clinicians to quickly diagnose PD but will also be useful to develop a better patient care strategy at an early stage of PD. In this study, various experiment are conducted to observe the most contributing speech feature to classify PD patients. The study have showed by using the Best-First feature selection approach the most optimal features from the PD dataset can be achieved. The efficacy of our approach with the optimal set of features has shown an improvement in classification with an accuracy of 92.19% that is better than the earliest reported accuracy of 86% [23] for an almost similar number of features.