Using reduced feature space (2D and 3D) based on entropic measures for detecting Parkinson’s disease through voice

Abstract

This research presents a proposal for the integration of different fields, including Parkinson’s disease (PD) detection, acoustic voice analysis, and signal processing. The proposal entails the development of two and three dimensional parsimonious models, predicated on feature spaces constituted by variants of Shannon’s permutation entropy and autocorrelation measures. These models elucidate the structural and informative nature of vocal signals in individuals with and without the disease (NPD). The reduced-dimensional feature spaces (2D and 3D) are novel and were used for the automatic classification of voices using support vector machines (SVM) with polynomial kernels and cross-validation, achieving average accuracy values between 0.82 and 0.88. Furthermore, the identification of homogeneous subgroups according to the coordinates in thespace of 2D characteristics represents significant progress. The variables under consideration are candidates for biomarkers of subtypes of speech disorders for Parkinson’s disease. The database used is freely accessible to facilitate reproducibility. The proposed approach is simple and precise and shows promise for the diagnosis and monitoring of PD through the effective use of samples of the vowel /a/ of just one second with a reduced feature space that could improve clinical workflows.