ECG beat classification using wavelets and SVM

Electrocardiogram (ECG) is one of the most important noninvasive tools for the diagnosis of cardiac arrhythmia. Automatic beat classification in ECG is a topic of continuing research. In this paper, automatic classification of 3 beat types - normal sinus rhythm, premature ventricular contraction and left bundle branch block is implemented. QRS detection is done using the Pan Tompkins algorithm. Wavelet decomposition using daubechies 4 wavelet is done. 25 features are extracted for each beat from wavelet analysis, namely - mean, variance, standard deviation, minimum and maximum of detail coefficients and of approximation coefficients. 3 RR interval features are also extracted for each beat. Beat classification is implemented by using OAO (One Against One) SVM (Support Vector Machine). 3 SVM´s are designed and final grouping is done by maximum voting. Novel method of feature selection is introduced. Feature selection for a particular SVM is done based on the beats to be classified by that SVM. ECG signals are obtained from the open source MIT-BIH cardiac arrhythmia database. 6355 beats (2036 LBB, 3865 N, 454 PVC) are used for testing the implementation. Accuracy of 98.46%, 98.47% and 99.92% are obtained for left bundle branch block, normal and premature ventricular contraction beats respectively.