DETECTION OF LUNG DISORDERS USING EMBEDDED AND WRAPPER FEATURE SELECTION METHODS

Abstract

Despite the advances in biomedical signal processing in recent years, the need for fast and highly accurate diagnostic systems for the detection of lung disorders continues. In the study, 150 normal and 444 abnormal lung sounds obtained by automatic detection of respiratory cycles from 94 different people by physical examination were used as a database. Then, 12 different feature extraction methods were applied in the time and frequency domain. Features were evaluated using embedded and wrapper feature selection methods. These methods are recursive feature elimination, adaptive structure learning, dependence-guided unsupervised feature selection, unsupervised feature selection with ordinal locality, feature selection via concave minimization, least absolute shrinkage, and selection operator feature selection methods. Features are classified by linear support vector machines, k nearest neighbor, decision trees, and naive Bayes classification methods. As a result, when the number of features is not limited, 97.3% accuracy is obtained when the recursive feature elimination method is used together with the k nearest neighbor classifier. In the case where the number of features is limited to three, the classification accuracy of 91.4% was achieved using the adaptive structure learning feature selection method and the decision trees.