Selection of myocardial electrogram features for use by implantable devices.

Abstract

Implantable devices that terminate ventricular tachycardia must be capable of correctly classifying heart rhythms to a high degree of reliability. We evaluated the relative discriminating power of several myocardial electrogram (ME) features in six human subjects by reducing the order of their corresponding feature spaces using three different optimization methods: 1) minimizing univariate Bayes error rates (univariate parametric), 2) maximizing the Kullback divergence (multivariate parametric), and 3) pruning classification trees (nonparametric). We found that although the composition of the optimal subspaces varied considerably from one subject to another, one frequency domain feature was common to most of the optimal subspaces.