Classification Algorithm For Pulmonary Hypertension Associated With Congenital Heart Disease Based On Fusion Of Heart Sound Signal Features

Congenital heart disease-associated pulmonary arterial hypertension(CHD-PAH)has a high mortality rate,which early symptom is not obvious and is not easily detected.Clinical confirmation for CHD-PAH requires interventional right heart catheterisation,which requires specialised equipment and is too risky and difficult to perform in mass screening.The timely and early diagnosis of CHD-PAH remains a challenge.Heart sounds are the sounds produced by the heart during systole and diastole and reflect the normal functioning of the heart valves and chambers.Based on the analysis of heart sound signals,a non-invasive,portable and timely diagnosis of CHD-PAH and early medical intervention is expected to reduce the delay in diagnosis due to various factors and improve the survival rate of CHD-PAH patients.Most of the traditional studies on heart sounds have focused on the classification of normal and abnormal heart sounds and have achieved good classification results.In this paper,the CHD-PAH heart sounds were analyzed and studied.By analyzing normal,CHD,and CHD-PAH heart sound signals,a kind of fusion feature was proposed,in which the time domain,frequency domain features,and deep learning features were included.The differences among above were studied and finally achieve the three-way classification of normal,CHD,and CHD-PAH heart sound signals.In this study,heart sound signals were selected from 483 patients,161 normal,161 with CHD,and 161 with CHD-PAH.Each patient’s heart sound signal was collected using the five-point sampling method of clinical auscultation.The raw heart sounds were first pre-processed,in which a double-threshold adaptive segmentation method was used to first segment the 20 s long signal into individual cardiac cycles.Each cardiac cycle and S2 component were then extracted for its corresponding time-domain features,and frequency-domain features.The 1D heart sound signal was then converted into a 2D feature map of size 34*34 by taking the de-discretized Power-Normalized Cepstral Coefficients for each heartbeat cycle.The depth features of the PCG signal are extracted by a convolutional neural network(CNN).Finally,the above features are combined into a fused feature vector that is used as input to the classifier.In order to classify heart sounds,k-NN(k-Nearest Neighbors)algorithm,Ada Boost(Adaptive Boosting),v-SVM,and XGBoost(Xtreme Gradient Boosting)were trained and then used as classifiers for detecting normal,CHD and CHD-PAH.The results show that XGBoost has best classification results compared with the other algorithms.Finally,considering that each case of 20-second heart sound samples can segment individual heart sounds to obtain between 20 and 33 heartbeat cycles,this means that multiple classification results can be obtained per case.The classification result is optimized to best by the majority voting algorithm,which avoids the effect of chance on the classification result for a single cycle.The first stage of this paper,the dichotomous classification of CHD and CHD-PAH is done,which is important to explore feature extraction and select classifiers during the study.Then the normal heart sound signals are introduced in the dataset.Finally,the fusion feature is input to classifiers to achieve triclassification of normal heart sounds,CHD and CHD-PAH.Using this method,an accuracy of 89.02% for dichotomous classification and an accuracy 88.61% for triple classification were achieved.Compared with existed studies,the method proposed in this paper has good prospects for application in non-invasive diagnosis of CHD-PAH,and a new pathway for the early diagnosis of CHD-PAH has been opened up.