Research On Computer-aided Diagnosis Of Common Congenital Heart Diseases

Facing the threat of human health killer-- heart disease, especially congenital heart diseases(CHDs), the diagnosis and treatment of cardiac diseases have become one of the current challenges in the medical field. In present, major treatment process can only be conducted in advanced medical institutions. However, for remote and poor areas such as in the countryside, it is difficult for patients to be successfully diagnosed due to the lack of advanced equipment. The aim of this study was to develop a novel method based on four kinds of clinical cardiac physiological signals acquired easily to achieve a simple diagnosis and effective method for diagnosing CHDs, which can be handily used in remote areas for patients to seek medical treatment.Clinical cardiac physiological signals used in this paper are: left ventricular end diastolic diameter(LVEDD), left ventricular ejection fraction(LVEF), B-type natriuretic peptide(BNP), cardiac defect size, based on the above four kinds of cardiac physiological signal classification and prediction models constructed a multi-step diagnostic systems. First, through data preprocessing, abnormal samples were removed from the original clinical data. And then, based on three kinds of classification, cardiac diseases are sub-striated into right categories. Lastly, according to the type of disease, neural network method is used to forecast for the sizes of heart defects.This paper addressed the following five key questions:(1) The multivariate binomial regression model was applied to de-noise and pretreat the obtained clinical data. Four different multivariate binomial regression models are used, including linear, pure quadratic, interaction and pure quadratic to fit the clinical data. These four methods are used to draw exception points comprehensively of the original clinic data.(2) Three main types of congenital heart diseases are investigated in this study, including the ventricular septal defect(VSD), atrial septal defect(ASD), patent ductus arteriosus(PDA), which were classified by multi-steps using classification methods, namely the Naive Bayes classifier, random forest classification and the support vector machine(SVM). Final classification was obtained by an integral consideration of each of the classification results obtained finally by majority voting. Firstly, distinguish the two diseases of VSD and PDA from ASD, and then distinguish VSD from PDA. The first step’s accuracy rate is above 85%, and the second stage classification is about 61%. The result confirms the fact that the physiological and pathological signal characteristics were very similar and almost occur concurrently in VSD and PDA.(3) Further, an integrated approach of BP neural network and RBF neural network is used to predict the size of the heart defects. Integrating the clinical cardiac physiological data on the learning and training of a network between the relational models with physiological signals to make prediction s based on high credibility. The mathematical relationship between clinical physiological data with models was built based on correlation analysis between the features.(4) A hierarchical diagnostic method of CHDs was proposed. The first was to use regression models for data preprocessing. Then, by using probabilistic and neural network models, an initial disease classification was made. Based on such initial classification result, neural network models was further used to predict the severity of heart disease, i.e., the heart defect size for various types of heart diseases. Finally, a completed process of diagnosis of congenital heart disease was achieved.(5) Solve problems arising from insufficient data for classification and model prediction. Such a solution is also applicable to neural network learning and probabilistic analysis, which normally require a large data set for accuracy and reliability.By solving the above five key issues of diagnosis of CHDs, we have obtained good classification and prediction results with clinical data. The developed methods may have good clinical application values.