Maximum Approximate Entropy for Normal and Pathological Voices Classification

The assessment of voice signals using non-linear features has proved to be valuable tool for the automatic detection of pathological voices. In this paper we propose a new approach based on the maximum approximate entropy estimator (ApEn(max)) and the r value at which it is achieved (r(max)). Through experiments with data from the MEEI voice disorders database, we evaluate the proficiency of these estimators as a function of the embedding dimension m. Using these features along with linear discriminant analysis and principal component analysis, we have achieved an accuracy of 94.6%, a sensibility of 98.2% and a specificity of 84.7%. We can conclude that the jointly use of ApEn(max), and r(max) can be used to discriminate between pathological and normal voices. Moreover, the discrimination capacity of a simple linear classifier can be increased using in conjunction the information brought by these estimators through several m values.