Um estudo de reconhecimento de sons pulmonares baseado em técnicas de inteligência computacional

Abstract

This work describes the use of Computational Intelligence techniques to classify pulmonary sounds from normal to adventitious. Normal sounds are auscultated in healthy subjects. Adventitious sounds are auscultated in subjects with lung disease, and are divided into two categories: continuous sounds (wheezes and rhonchus) and discontinuous sounds (crackles). Each is related to pulmonary dysfunctions, making it important to classify these sounds to support clinical diagnosis. In addition, pulmonary sounds are non-stationary signals, which makes them difficult to analyze and hard to distinguish when using traditional auscultation methods such as a stethoscope. Thus, the development of a technique to classify these sounds may aid professionals in performing clinical diagnosis. This study proposes the development of a pulmonary sound classifier using higher-order statistics (HOS) to extract features, Genetic Algorithms (GA) and Linear Discriminant Analysis to reduce dimensionality and Decision Trees, k-Nearest Neighbor, Bayesian Classifier and Support Vector Machines in order to classify pulmonary sound events. The pulmonary sound classes are: normal, fine crackles, coarse crackles, monophonic wheezes and polyphonic wheezes. The results obtained in this work revealed that the divide-and-conquer approach, employing k-Nearest Neighbor and Bayesian classifier, is most appropriate for the purpose of pulmonary sound classification, given that this approach achieved better performance in comparison with the use of only one classifier. The mean validation classification accuracy obtained by the divide-and-conquer approach was of 91.1%, which shows the efficiency of the proposed method.