Özet:

In this study, detailed feature analysis was performed on the pathological and healthy subjects via single-channel common lung sounds. The database obtained from the automatic detection of 594 respiratory cycles of normal, rhonchi, fine crackle, and coarse crackle sounds obtained from 94 people was used. Then, zero-crossing rate, energy, the entropy of energy, spectral centroid, the spread of a frame, spectral entropy, spectral flux, spectral roll-off, Mel-frequency cepstral coefficients, harmonic ratio, the fundamental frequency of a window, and chroma vector feature extraction methods have been applied to the database. Infinite latent feature selection, infinite feature selection, eigenvector centrality, minimum redundancy maximum relevance, relief, mutual information, laplacian score, multi-cluster, fisher, unsupervised discriminative, local learning-based clustering, correlation-based feature selection methods are used in the training phase. Support vector machine, k nearest neighbors, naive bayes, and decision trees algorithms were used for classification. As a result, 97.5% classification accuracy was achieved by using the k nearest neighbor classifier and the multi-cluster feature selection method when the number of features is not limited. When the number of features is limited to 3, 91.6% classification accuracy is achieved by using the k nearest neighbor classifier and the eigenvector centrality or infinite feature selection methods.

Anahtar Kelimeler: