Feature Selection and Comparison of Classifiers for Reinke’s Edema Identification

This study focus in the detection of Reinke’s edema based on vowel recordings. This benign disorder is characterized as edematous changes on the vocal folds, altering the produced voice. In order to assist in the diagnosis, we analyzed the effectiveness of Wavelet-Packet decomposition with different families (Haar and Daubechies) up to the seventh level of sustained vowel signals and compared the performance of support vector machine and multilayer perceptron classifiers to identify this disorder. One of the contributions of this work is to provide a filter method for selecting attributes from the Wavelet decomposition leaves. We show that less than 50% of the leaves are relevant for classification, but the ones from the initial, central, and final portions of the decomposition proved to be important, indicating that components from all frequency ranges help in the detection of Reinke’s edema. The Daubechies db5 family showed the highest overall accuracy in the tests, and the db2 family proved to be more consistent. It can be seen that the support vector machine, in addition to performing comparably to the multilayer perceptron, stands out for its significantly shorter training times, contributing to the practical viability of this approach in diagnosing Reinke’s edema.