Respiratory Adventitious Sound Classification Using Third-order Difference Plots

This study aims to classify abnormal respiratory sounds such as crackles and wheezes using third order difference plots. After obtaining difference plots from normal respiratory sound segments and segments with crackles and wheezes, the eigenvalues of the covariance matrices were calculated and used as features for classification. Binary classifications were performed using the decision tree algorithm, where the performances were evaluated through 10-fold cross-validation. Generally, higher performances were observed in distinguishing crackles and in the inspiration phase. The highest performance was achieved in crackle vs. normal classification in inspiration with an accuracy rate of 82%. This performance, obtained without additional features, suggests that difference plots are suitable tools for the classification of normal and abnormal respiratory sounds. In the future, it is aimed to test other classification algorithms, to improve the methodology with preprocessing steps applied on the signals and with new discriminative features calculated from the plots, and finally, to develop the three-class classifier for the form of usage that will benefit the clinical practice.