Methodology for Automatic Classification of Adventitious Lung Sounds

摘要

This paper describes the development of a methodology for adventitious lung sounds classification using discrete wavelet transform (DWT) and a classifier based in a radial basis function (RBF) neural network. The proposed algorithm classifies abnormal sounds into four groups: normal, continuous and discontinuous adventitious lung sounds, notifying the occurrence of both adventitious sounds. In the processing stage, the respiratory cycle is decomposed up to its 10th decomposition level and the energy for each level is calculated. The resulting curves show different draws for each kind of adventitious sound. Therefore, those curves are used as data source for a RBF neural network, which acts as an automatic classifier. For the computation of the results, ten different mother wavelets was tested and a hundred neural networks was trained for each mother wavelet, in a total of a thousand neural networks trained. The results achieved a maximum performance of 92.36%, showing that the energy versus decomposition level curves may be successfully used to classify the adventitious sounds.