Is it possible to distinguish different types of ECG-holter beats based solely on features obtained from windowed QRS complex?

Abstract

The main focus of this paper is to investigate the possibility to distinguish among different classes of beats, as provided by ANSI/AAMI EC57:1998 standard, from the ECG holter recordings. We compare the performance of an ensemble classifier based on three classifiers on distinguishing ECG beats from holter recordings characterized by two distinct sets of features. The first feature set is one relying upon the "classical" time interval measurements of QRS complex and T-wave. The second one tries to describe the beat using means as simple as possible resulting in a description of the QRS complex in terms of "easy-to-compute" statistical moments; hermite coefficients and Karhunen Loeve coefficients. The results of the ensemble classifier consisting of three different classifiers - namely a k-NN classifier, a Back propagation Neural Network and a Support Vector classifier-are as general as possible by using global training/ testing approach that uses one half of the recordings from the MIT-BIH database for training and the other half for testing. Results of the classifier are computed using sensitivity (Se) and specificity (Sp) for both feature sets. The best results achieved during the experiments were those using the "classical" feature set and the ensemble classifier. The specificity for detection of normal beats was 74.26% and sensitivities were 68.19%, 45.73%, 35.19%, 48.70% for ventricular, bundle branch blocks, supraventricular, and fusion beats respectively. The results achieved on the "easy-to-compute" approach are comparable to those from "classical" approach when dealing with the detection of ventricular beats with specificity 74.73% and sensitivity 59.97% - but they have performed much worse when trying to detect the other classes such as supraventricular, fusion or bundle branch block beats.