Feature and Classifier Selection for Respiratory Sound Classification

In this study, Mel Frequency Cepstal Coefficients, Autoregressive(AR) parameters and their combination are com-pared as features in classifiers for recognizing pathological and healthy subjects. Results show that AR parameters outperform both MFCCs and combined features. For fast and efficient classification, AR parameters of respiratory sounds are studied to characterize lung sounds for diagnosis of pathological subjects. Various time domain, frequency domain and time-frequency domain features are added to the feature set. After feature extraction step, feature selection based on feature importance scores and SVM-RFE are used as feature selection step. Experiments are conducted on a dataset of 30 subjects and several machine learning algorithms are used as for classification. For optimum computation time and classification accuracy, we propose a method based on random forests. The proposed method achieves an accuracy of 93.3 % for 30 subjects.