Application Of Kernel Independent Component Analysis In Arrhythmia Pattern Classification

ECG signal is a synthetic reflection of the heart electricity on body surface. It has important significance to the diagnosis of heart disease by clinical ECG examination. Recognizing arrhythmia on its early stage is very important for diagnosing and forecasting the state of illness, and it is also critical for the further treatment of the patients. Because the signal is a typical non-stationary random signal, it's very complex and nonlinear, which increases the degree of difficulty on analysis, and the automatic diagnosed effect can not achieve to be satisfied. So the researchers are still persistent to improve the traditional methods and explore new resolution.According the features of ECG, As the viewed from extracting nonlinear features ,This thesis proposes a new method based on Kernel Independent Component Analysis (KICA) as the feature extraction method of ECG, The discrete wavelet coefficients and time domain features are also added to build a multi-domain feature vector. Then a feature selection based on mutual information max-dependency and max-relevance is applied to select the best feature subset for different ECG representation.We choose Support Vector Machine (SVM) as the classifier to accomplish the classification. SVM is a pattern classification method based on structural risk minimization principle. Because of its strong learning ability and generalization ability SVM has been applied to many areas. In our experiments, we design multi-class classifier by employing error correcting output coding and SVM(ECOC-SVM), and use the AUC(Area Under the ROC Curve) as the function to evaluate classification result. Then obtain an integrated ECG signal recognition system and a better measure of classification performance.The method is used to recognize five types of arrhythmia heartbeats from the data provided by MIT/BIH arrhythmia database. Each kind of feature extraction methods proposed in the thesis are tested in experiments and compared, Finally, the result show that the method proposed can present the characteristics of different ECG type immensely, And increases the predicted classification accuracy with good generalization performance of support vector machine.At last, We design a arrhythmia pattern classification system by the matlab Graphical User Interface(GUI), and I hope it is helpful for studying and developing ECG automatic analysis technology.