QRS classification using adaptive Hermite decomposition and radial basis function network

The accurate QRS classification is essential for enhancing the performance of automated ECG interpretation system. The correct selection, of features is very important to minimize the dimension of the feature space and maximize the distance between different classes of QRS. In this study, the QRS complexes are decomposed on a set of Hermite functions and the coefficients attained are served as the features. The Hermite functions are orthonormal and have strong morphological representation ability for QRS complexes; thus each feature extracted has independent information and the signal can be represented with a low number of coefficient. An Adaptive Hermite Decomposition (AHD) method is employed to perform the decomposition. Then, a Radial Basis Function Network (RBFN) is used to classify normal and PVC patterns in the features space. Because the patterns are relatively concentrated, the choice of RBFN is reasonable. The results show that the RBFN can provide a satisfactory classification with a few training epochs. Based on MIT/BIH database annotations, 600 pairs of normal and abnormal QRS complexes (PVC beats) were extracted from 12 files to train and test the classifier developed using AHD and RBFN