A COMPARISON OF SIMILARITY MEASURES FOR CLUSTERING OF QRS COMPLEXES

Abstract

Similarity or distance measures play important role in the performance of algorithms for ECG clustering problems. This paper compares four similarity measures such as the city block (L1-norm), Euclidean (L2-norm), normalized correlation coefficient, and simplified grey relational grade for clustering of QRS complexes. Performances of the measures include classification accuracy, threshold value selection, noise robustness, execution time, and the capability of automated selection of templates. The clustering algorithm used is the so-called two-step unsupervised method. The best out of the 10 independent runs of the clustering algorithm with randomly selected initial template beat for each run is used to compare the performances of each similarity measure. To investigate the capability of automated selection of templates for ECG classification algorithms, we use the cluster centers generated by the clustering algorithm with various measures as templates. Four sets of templates are obtained, each set for a measure. And the four sets of templates are used in the k-nearest neighbor classification method to evaluate the performance of the templates. Tested with MIT/BIH arrhythmia data, we observe that the simplified grey relational grade outperforms the other measures in classification accuracy, threshold value selection, noise robustness, and the capability of automated selection of templates.