Identificação de estertores em sons respiratórios utilizando transformada wavelet e análise de discriminante linear

Abstract

Crackles are adventitious and discontinuous breath sounds that occur in lung diseases. Time domain parameters classify the crackles as fine, medium, and coarse, and may have positive or negative polarity. This work investigates methods and tools to characterize and classify crackles. Samples of breath sounds containing crackles were normalized and resampled at 8 kHz. Several experiments using the discrete wavelet transform (DWT), linear discriminant analysis (LDA), and k-NN have been performed, and evaluated with ROC analysis. A pattern recognition system was implemented with DWT, LDA and k-NN to classify fine and coarse crackles, and normal breath sounds. The experiment with different signal border extension methods during DWT decomposition showed the influence on the results of the characterization. The results indicate that the methods ZPD, SP0, SYMH, SYMW, ASYMH, PPD and PER are recommended, while SP1 and ASYMW methods are not recommended for the decomposition and characterization of crackles because they generate different characteristics in the higher subbands. Another experiment showed that the characterization of crackles using DWT can be made using certain decomposition subbands (D3, D4, and D5 with signal sampled at 8 kHz), thus reducing the computational effort. Another classification system implemented using LDA and DWT showed that crackles can be classified by their polarity indicating a high degree of accuracy (AUC rate up to 0.9943 for Symlet 19). Two experiments were conducted for mother-wavelet selection that best characterizes crackles. The first one quantitatively evaluated the similarity between the crackle and several mother-wavelets using Pearson's correlation coefficient. The mother-wavelet that resulted a strong correlation with the crackles, being most indicated for use were: Reverse Biorthogonal 3.7, 5.5 Biorthogonal Reverse, Reverse Biorthogonal 3.5, Daubechies 5, Symlet 5, Daubechies 6, 7, and Symlet Daubechies 7. The second experiment selected mother-wavelets by the power concentration in subbands. Previous trials already shown that the energy of the crackles decomposed by DWT are concentrated in a few subbands, so mothers-wavelet that concentrate larger percentage of the energy in a specific subband were selected, which were Daubechies 7, Symlet 7, Coiflet 3 and Symlet 12. The final experiment performed was a combination of mother-wavelets to improve the separability of crackles and normal breath sounds. The experiment showed that a classification system using DWT, LDA, and a linear classifier may totally separate the two classes (AUC ratio = 1) when the combination of mother-wavelets to generate the feature vector of the signals is used.