Clustering Of Ecg Cycles

Abstract

The study is focused on a design of a reliable approach for ECG cycles clustering. It would be helpful for automatic assessment of various pathological patterns in ECG. Proposed method was tested and tuned on real data from ambulatory ECG database. The algorithm comprises ECG preprocessing, adjustment of R-peak positions available in database, creation of a template cycle, computation of features mainly representing correlation between particular cycles and the template, and, clustering of cycles within ECG via k-means. The appropriate number of clusters is derived via analysis of silhouette values. Resulting success of the algorithm in comparison with available manual scoring is: Sensitivity = 0.55 and Specificity=0.94.