Selection of voice parameters for Parkinson´s disease prediction from collected mobile data

Voice disorders, which can help in the diagnosis of Parkinson's disease (PD), can be measured with acoustic tools. In this work, demographic data and vocal phonation records /a/ from the available mPower database were analyzed to identify PD patients. A parsimonious model was then found that achieved a reduction from 62 to 5 phonation characteristics, which were considered in addition to gender and age. Multilayer Perceptron (MLP) and Logistic Regression (LR) neural networks were used to obtain a model with high prediction capacity (area under receiver operating characteristic curve, AUC-ROC, over 0.82). This work contributes to the monitoring of EP patients from the recording of a few phonation features collected by means of a mobile phone.