Sayısal Stetoskop ile Elde Edilen Kalp Ses (Fonokardiyogram) Sinyallerinin Bölütlenmesi ve Sınıflandırılması

Abstract

The main purpose of this work is to develop a decision support mechanism that will contribute to the early diagnosis of extraordinary heart diseases by the help of a digital stethoscope, especially for primary care physicians. It is an important goal and field of study to create an automatic system that gives all physicians the ability to auscultate in an experienced cardiologist's profession by closing the differences of knowledge and experience among the physicians by providing a system to hear and understand the heart diseases that the physicians can not easily hear. The classification of heart diseases depends on the correct identifications of S1 and S2 segments. Without the ECG reference signal, segmentation methods become naturally more complicated. In the hospital environment, the heart sounds collected from the patients by the stethoscope carry many environmental sounds such as ambient noise, speech, wheezing, and friction. Besides, depending on the heart condition, noise like murmur is also included in these heart sounds. Discrete Wavelet Transform (DWT) and Mel-Frequency Cepstral Coefficient (MFCC) are used as a hybrid solution for the filtering of the noise content of basic heart sounds. In order to determine S1-S2 locations, heart rate and systolic time intervals are determined using signal autocorrelation. As a result of this proposed algorithm, S1 and S2 sounds are detected with 98.19% precision and 98.52% recall for normal heart sounds, while S1 and S2 are detected with precision of 94.31% and recall of 96.92% for abnormal heart sounds. Ten different discriminating features are derived for each sound interval from the segmented heart sounds. With supervised classification methods, identification of Aortic Stenosis, Mitral Regurgitation, Aortic Regurgitation, Mitral Stenosis and Normal heart sound signals are performed. In this study, the Support Vector Machines, K-Nearest Neighbor, Artificial Neural Networks and Decision Trees supervised classification methods are compared with respect to the classification performance over these diseases. Classification performance is separately reported for Support Vector Machines, K-Nearest Neighbor, Artificial Neural Networks and Decision Trees of Aortic Stenosis, Mitral Regurgitation, Aortic Regurgitation, Mitral Stenosis and Normal heart sound signals. From these results it can be seen that different classifiers can perform well for different problems.