Feature selection based on binary particle swarm optimisation and neural networks for pathological voice detection

In this work, 52 Haralick texture features, extracted from two-dimensional wavelet coefficients of speech signals from recurrence plots (RPs) pathologies are used for pathological voice discrimination. Here, three pathologies are considered for analysis: vocal fold paralysis, edema and nodules. For feature selection, a binary particle swarm optimisation (PSO) algorithm using multilayer perceptron (MLP) neural network with cross validation is employed. The adopted fitness function is based on the maxima average accuracy rate. Statistical tests for individual measures were made and their results show statistical significance for several employed measures. The measures were combined and the more relevant ones based on the highest accuracy were selected by the PSO. The comparison with and without PSO by applying the statistical test of mean difference showed that the PSO use increased the accuracy rates. Furthermore, the PSO use reduced the amount of features for almost half of all initially used.