Session S92.1

Analysis of Multi Domain Features for ECG Classification

M Llamedo*, JP Martínez

Universidad Tecnológica Nacional
Buenos Aires, Argentina

Automatic ECG classification is a well-known problem for which a number of models have been proposed in the last two decades. The aim of this work is to evaluate the utility and robustness of single and multi-lead features, like intervals, energies, amplitudes and directions for ECG classification. The MIT-BIH Arrhythmia database was used together with AAMI-recommended class labelling. In this study, we have considered classical features based in the heart rhythm as well as features obtained from different scales of the wavelet transform (WT) like amplitudes, energy and direction of the WT of the QRS complex, the P and T waves. We used a quadratic spline wavelet prototype, which has been applied with good results for ECG delineation. The MIT-BIH database (DB) was divided into 2 datasets, one for model selection and training (DS1) and the other for testing (DS2). The best model was selected with a sequential floating feature selection algorithm, with k-fold cross validation for selecting the best features for a quadratic classifier in DS1. The best performing model in DS1 was used for the final performance calculation in DS2. The features which obtained higher score were the RR interval, the width of the QRS complex, features related to the circularity of the QRS loop at scale 4 and 5, the area of the WT signal at scale 2 and others related to the energy previous to the QRS complex at scales 3 and 4. The best model was of 11 features and obtained (in DS2) for the normal class a sensitivity (Se) of 60% and positive predictability (+P) of 98%, for supraventricular beats a Se of 83% and a +P of 27%, for ventricular beats a Se of 83% and a +P of 28%. For both the fusion and unclassifiable beats the classifier showed bad results because of being poorly represented in this DB. The total accuracy achieved was of 70%. The methodology presented allowed to evaluate the utility of features in the context of ECG classification. As a result, a baseline model is suggested for future improvements in the feature space and in the classifier stage. This methodology is particularly useful when used after a wavelet based delineator, since the features can easily be derived from the wavelet coefficients used for delineation.

(Abstract Control Number: 31)