Research On ECG Beat Classification Algorithm Based On Nonlinear Feature Extraction

Heart disease is one of the highest morbidity and mortality of disease on medicine. The electrocardiogram (ECG) is an important technique for heart disease diagnosis. Therefore, the analysis of the ECG has been widely researched, which mainly involves two aspects:ECG signal feature extraction and classification. ECG signal classification is an important diagnosis tool wherein feature extraction plays a crucial function.Several important issues exist in the ECG beat classification, which, if suitably addressed, may lead to development of more robust and efficient recognizers. Some of these issues include feature extraction, choice of classification approach, and optimization. Extraction of effective feature vector is an important module in arrhythmia classification, in order to be able to more fully reflect the characteristics of the ECG signal, this thesis puts forward a kind of nonlinear feature extraction method based on wavelet packet decomposition (WPD), empirical mode decomposition (EMD) and approximate entropy (ApEn). The proposed method first uses db6 wavelet to decompose ECG signals into different frequency bands by WPD and then calculates the ApEn of each wavelet packet coefficient as a feature vector, then uses EMD to decompose ECG signals into a finite number of intrinsic mode functions (IMFs), calculates the ApEn of IMFs as another feature. The two features are regarded as a feature vector.The probabilistic neural network (PNN) and support vector machine (SVM) are used for ECG beat classification. The most important disadvantage of PNN is its difficulty in selecting an appropriate spread value. Classification performance depends heavily on the selection of an appropriate spread value. A drawback of SVM is its choice of the kernel function and parameters. RBF is generally applied because it can produces higher accuracy rates classifying multi-dimensional data compared to a linear kernel function. RBF also presents fewer parameters with which to set than a polynomial kernel function. Therefore, RBF is an effective option for producing the kernel function. The choice of value for C and a greatly affects the classification outcome. The particle swarm optimization algorithm is used to optimize parameters of the PNN and SVM. The performance of the SVM classifier is slightly superior to that of the PNN classifier with 98.6% accuracy.Finally, the thesis adopts the Fluke prosim 2 human vital signs simulator as the signal source to build experimental platform for ECG signal acquisition. Then the features are extracted and classifications are made based on the method for validation. The results show that t the classification accuracy and the overall degree of sensitivity, specificity and positive predictive value of SVM is slightly higher than the PNN, and has achieved higher specificity of 98.5%. In the case of a small amount of samples, the results are satisfactory.