Benchmarking dynamic time warping on nearest neighbor classification of electrocardiograms

Abstract

The human cardiovascular system is a complicated structure that has been the focus of research in many different domains, such as medicine, biology, as well as computer science. Due to the complexity of the heart, even nowadays some of the most common disorders are still hard to identify. In this paper, we map each ECG to a time series or set of time series and explore the applicability of two common time series similarity matching methods, namely, DTW and cDTW, to the problem of ECG classification. We benchmark the two methods on four different datasets in terms of accuracy. In addition, we explore their predictive performance when various ECG channels are taken into account. The latter is performed using a dataset taken from Physiobank. Our findings suggest that different ECG channels are more appropriate for different cardiovascular malfunctions.