Selección de Medidas de Disfonía para la Identificación de Enfermos de Parkinson [Not available in English]

The non-invasive nature of speech testing in patients with Parkinson's disease (PD) makes it an important tool for the detection and monitoring of this disease. In this work, 297 measures of dysphonia obtained from 108 records of people with and without the disease are used. These measures were classified according to four groups, where each one characterizes different pathologies in the voice: two groups are associated with signal periodicity, one group is associated with noise problems due to incomplete closure of the vocal cords, while the last group is related with difficulties in positioning articulators in vocal tract to achieve phonation of the /a/. A pre-selection of 35 measures was made according to the multicollinearity among the total group and variables considered relevant in the bibliography. The implementation of penalized estimation algorithms for feature selection in logistic regression models, made it possible to obtain models whose average predictive capacities exceeded 80%. The results of this study allow obtaining a hierarchical classification of dysphonia measures, important for the detection and monitoring of PD with non-invasive techniques.