A comparison of the ECG classification performance of different feature sets

This study investigates the automatic classification of the Frank lead ECG into different disease categories. A comparison of the performance of a number of different feature sets is presented. The feature sets considered include wavelet-based features, standard cardiology features, and features taken directly from time-domain samples of the EGG. The classification performance of each feature set was optimised using automatic feature selection and choosing the best classifier model from linear, quadratic and logistic discriminants. The ECG database used contains 500 cases classed into seven categories with 100% confidence. Using multiple runs of ten-fold cross-validation, the overall seven-way accuracy of different feature sets and classifier model combinations ranged between 60% and 75%. The best performing classifier used linear discriminants processing selected time-domain features. This is also found to be the simplest and fastest classifier to implement.