Clinical Correlate of Dysphonia Measures in Patients with Parkinson’s Disease

Abstract

In patients with Parkinson's disease (PD), the impairment of neurological control is frequently observed to affect the normal functioning of the muscles involved in speech production. In this study we propose to analyze a speech corpus from subjects with and without PD, with the aim of quantifying speech alterations. The data corpus was built in 2019 in collaboration with public hospitals and universities of Argentina, it contains 55 audio samples from patients with PD and 71 subjects without PD. The analysis was conducted from four perspectives: a) perceptual speech therapy evaluation of voice quality; b) estimation of acoustic disturbance through the integrated perturbance index; c) otorhinolaryngological clinical evaluation to identify structural and functional alterations; and d) acoustic analysis of the sustained phonation signal of the vowel /a/. In the latter, 339 measures of dysphonia were estimated, which were then reduced to 11 parameters using statistical feature selection methods. These parameters achieved a performance greater than 80% in the automatic binary classification task (PD or non-PD). In the otorhinolaryngological clinical evaluation analysis, the clinical findings were classified into four categories based on the presence or absence of the alteration: gastroesophageal reflux; tremor during phonation and rest; alterations in glottic closure; and other alterations. We estimated the degree of correlation between the analyses from these four perspectives. The integrated perturbance index is more relevant than the perceptual analysis of the voice quality, however, no correlation is observed with the clinical evaluation. Nonparametric statistical tests were used to assess mean differences between the 11 measures of dysphonia and each of the four categories of clinical findings identified. Significant differences were observed for four of the 11 selected dysphonia measures. These results encourage us to continue our analysis of acoustic parameters. We are also beginning to build a new data corpus, in which other measures of interest will be added.