Classification of heart valvular disorders using Time-frequency analysis of phonocardiogram signal and machine learning methods
Abstract:
During the last few decades, the prevalence of heart diseases in the world has been increasing due to lifestyle changes and other unknown factors. Heart valve disorders (HVDs) have a high mortality rate compared to other cardiovascular diseases (CVDs) and are considered as the main cause of cardiovascular diseases (CVDs). Therefore, their prediction and diagnosis are of great importance in the field of medicine, because accurate diagnosis in the early stages followed by appropriate treatment helps to reduce the death rate caused by CVD. The phonocardiogram signal is a cost-effective and non-invasive cardiac auscultation tool that contains heart sounds in a cardiac cycle and allows for better interpretation of heart valve function. Skilled cardiologists usually analyze these sounds, which are caused by muscle contractions and the closing of heart valves. However, this process can be affected by factors such as environmental noise, limitations in the hearing frequency range and physician expertise - lack of proper training to record vital information from the cardiac signal, and finally the time-consuming process of visual screening of the PCG signal. Therefore, an automatic diagnosis process based on time-frequency analysis of the phonocardiogram signal using machine learning algorithms will be necessary for doctors to diagnose heart valvular disorders. Also for smart healthcare applications, integrating signal processing with machine learning techniques is a new research trend in heart sound analysis studies to detect cardiac abnormalities. Therefore, it is important to develop new methods for processing time-frequency phonocardiogram signals and classifying valvular heart disorders using machine learning algorithms. In this thesis, the combination of time-frequency phonocardiogram signal processing and machine learning algorithms is used to diagnose people with valvular heart disorders with high accuracy and precision. In this research, using the Yasin five-class database, time-frequency decomposition methods based on three approaches of discrete wavelet transform (DWT), empirical mode decomposition (EMD) and empirical wavelet transform (EWT) have been used. In the proposed algorithms in all three approaches, common statistical features in this field were extracted from the decomposed components. The relieff feature selection method was used for optimization and classification with a smaller number of features, then svm, knn, ensemble, naïve bayes classifiers have been used for classification.
Finally, the results obtained from the proposed methods for classifying five classes of heart valve disorders show that, for the EWT+ Relief method, 98. 68% accuracy was obtained using 180 features and the Ensemble classifier, 98. 48% accuracy was obtained using 100 features and the Ensemble classifier, and 98. 12% accuracy was obtained using 20 features with the SVM classifier. In the EMD+ Relief method, 98. 88% accuracy was obtained using 72 features with the Ensemble classifier, 99. 04% accuracy was obtained using 50 features with the Ensemble classifier, and 99. 20% accuracy was obtained with 20 features with the KNN classifier. In the DWT+ Relief method, 99. 44% accuracy was obtained using 144 features with the Ensemble classifier, 99. 56% accuracy using 100 features with the knn classifier, and 99. 52% accuracy with 20 features with the knn classifier. As can be seen from the results, the DWT+ Relief method is superior to other methods. These values show promising results for improving the process of diagnosing valvular heart disorders. Using the DWT feature extraction algorithm along with the selection of the Relief feature can lead to more accuracy in diagnosing valvular heart diseases and can be presented as a practical proposed method in the medical field with a variable number of features to the relevant doctors to be used for better diagnosis of valvular heart diseases.