A Wavelet Based Classifier of the Electrocardiogram

The classification of the electrocardiogram (ECG) is essentially a pattern recognition task where successful classification relies on efficient representation of the diagnostically significant shapes of the ECG. This study investigates the use of selected wavelet coefficient values to represent the diagnostic information. The coefficients are used as inputs to a linear discriminant classifier. A two-stage feature selection scheme has been implemented to determine diagnostically useful sets of coefficients. The first stage selected a set of candidate coefficients by choosing coefficients with the highest magnitude. The second stage chose a subset of the candidate coefficients that maximised the classification accuracy by using a forward selection process. Our method was verified using a 100% accurate database of 500 ECG records classified into 7 classes. Using an initial set of 1238 coefficients 26 wavelet coefficients were chosen and the classifier achieved an overall accuracy of 73.4%. This compares well to the CSE “panel of cardiologists” result of 74.8%