ABSTRACT
Arrhythmia is the most fatal for human being among all cardiovascular diseases. Early detection of arrhythmia beats, from long term ECG record, is helpful to start treatment and saving life of patients. In this work, we presented a patient-adaptive scheme to discriminate normal and three classes of arrhythmia beats from ECG signal. Instead of conventional features, the proposed method uses a kernel based modeling technique of the ECG beats and the model coefficients are used as the features to characterize different types of beats. In this semi automatic scheme, a global training set is combined with a local learning set to form a patient adaptive training set to develop a patient specific classifier model. The results are validated on MIT-BIH arrhythmia database and the performance of the proposed technique is validated by three classifiers namely, support vector machine (SVM), vector valued regularized kernel function approximation (VVRKFA) technique and k-nearest neighbour (KNN) classifiers. Experimental results indicate that the proposed patient adaptive classification scheme increases the global accuracy by 12 to 16% than that of the accuracy obtained without using patient specific beats to global training set. The highest average accuracy obtained using this method is 96.63%, which is comparable and even better than most of the works available in the literature.
