Conference Program

A Predictive Model for Speech Rehabilitation for Patients with Parkinson’s Disease (310051)

*Ismail EL Moudden, Eastern Virginia Medical School Sentara Healthcare Analytics and Delivery Science Institute 
Jiangtao Luo, School of Health Professions, Eastern Virginia Medical School 
Mounir Ouzir, Higher Institute of Nursing Professions and Health Techniques, ISPITS Beni Mellal, Regional Hospital 
Mohan Dev Pant, School of Health Professions, Eastern Virginia Medical School 

Keywords: Dysphonia Features, Parkinson’s disease, Machine Learning, Feature Selection, Class Predection

The primary aim of this study is to develop a prediction model to accurately predict the need for speech rehabilitation for patients with Parkinson’s disease. After using standard data preprocessing steps, the identification and selection of most relevant features contributing to speech recognition and rehabilitation was achieved by examining F-values of one-way ANOVA procedure. To develop an advanced prediction model for speech rehabilitation, we used classifiers such as logistic regression (LR), support vector machine (SVM), K-nearest neighbors (KNN), Adaboost, and artificial neural network (ANN) on the Lee Silverman voice treatment (LSVT) dataset, a publicly available dataset on Parkinson’s disease. Feature importance was estimated and evaluated by using feature importance ranking measure (FIRM) [Zien et al. 2009]. The results indicated that the number of features were reduced from p = 309 to k = 7 for the LSVT dataset. Using only k = 7 dysphonia features, the classification accuracy was around 91% for all five classifiers. Additionally, 5 of the k = 7 features contributed toward 80% accuracy of the Adaboost-based prediction model. In conclusion, this proposed dimensionality reduction approach has a potential to substantially reduce the number of features and increase prediction accuracy of the prediction model to classify patients into Parkinson’s disease patients or healthy speakers.