Heart Abnormality Classification by Phonocardiogram Analysis Using Fusion in Feature and Decision Levels

Useful information about the act of the heart valves can be provided by the audio signal. This leads to a fast, inexpensive and non-invasive heart diagnosis. Due to human auscultator limitation and like of stationary of phonocardiogram signals (PCG), diagnosing through the heard sounds by a stethoscope needs a lot of experiences. Therefore, an automatic system to classify biomedical signal PCG is required. ECG signal is required to accurately segment the heart sound signal. But, acquiring ECG is generally expensive and time consuming. This study proposes a segmentation free system to classify the PCG signals. For extraction of appropriate features from the PCG signals, various methods such as non-uniform filter banks based on maximum entropy, wavelet transform (WT), Mel Frequency Cepstral Coefficient (MFCC) and fractal features are used. Features are given to three classifiers: ML (Maximum Likelihood), KNN (K-Nearest Neighbor) and SVM (Support Vector Machine). The Dempster-Shafer decision Fusion rule is utilized for final decision making. The experiments were performed on PhysioNet/Computing data sets to evaluate the performance of various methods. Sensitivity, Specificity and kappa coefficients were obtained from all six data sets. It is found that the proposed method has a better performance compared to other methods.