Session S34.4

Support Vector Machine Based Arrhythmia Classification Using Morphological Descriptors

R Besrour*, Z Lachiri, N Ellouze

LSTS\ENIT
Tunis, Tunisia

The electrocardiogram is the most important biosignal used by cardiologists for diagnostic purposes. The ECG signal provides key information about the electrical activity of the heart. Automated classification of heartbeats has been previously reported by other investigators using a variety of features to represent the ECG and a number of classification methods. This paper introduces a new method of heartbeat classification based on the support vector machine classifier using morphological descriptors. The results of numerical experiments on the recognition of normal sinus beat and 11 other types of arrhythmias are presented and discussed. The proposed approach is validated in the MIT-BIH Arrhythmia Database. Every heart beat is characterized by 10-elements vector representing information of the amplitude, the area, specific interval durations and slopes. From our database, we extract a training and a test bases. For this, we use a cross validation method: 2/3rd of randomly selected data was used as the training set and the remaining 1/3rd was used as the testing set. We repeat the experiment several times and we average the results. We have used Gaussian radial kernel with σ=0.5 and a polynomial kernel with d=1. The parameter C used in experiments was set to C=100. To evaluate the performance of the classifier, four statistical indices are used: sensitivity, specificity, positive predictive value and negative predictive value. The average of classification indices obtained with polynomial kernel were a sensitivity Se of 86.26%, a specificity about 75.10%, positive predictive value 78.63% and negative predictive value 83.94%. Using the Gaussian radial basis, we increased the sensitivity to 90.70%, the specificity to 72.26%, the positive predictive value 82.20% and negative predictive value 89.68%.

(Abstract Control Number: 60)