Extraction of low-dimensional features for single-channel common lung sound classification.

Abstract

In this study, feature extraction methods used in the classification of single-channel lung sounds obtained by automatic identification of respiratory cycles were examined in detail in order to extract distinctive features at the lowest size. In this way, it will be possible to design a system for the detection of lung diseases, completely autonomously. In the study, automatic separation and classification of 400 respiratory cycles were performed from the single-channel common lung sounds obtained from 94 people. Leave one out cross validation (LOOCV) was used for the calibration and validation of the classification model. The Mel frequency cepstrum coefficients (MFCC), time domain features, frequency domain features, and linear predictive coding (LPC) were used for classification. The performance of the features was tested using linear discriminant analysis (LDA), k-nearest neighbors (k-NN), support vector machines (SVM), and naive Bayes (NB) classification algorithms. The success of combinations of features was explored and enhanced using the sequential forward selection (SFS). As a result, the best accuracy (90.14% in the training set and 90.63% in the test set) was acquired using the k-NN for the triple combination, which included the standard deviation of LPC and the standard deviation and the mean of MFCC.