Comparison of different methods of measuring similarity in physiologic time series

Abstract

Searching for similarity between time series plays an important role when large amounts of information need to be clustered to integrate intelligent supported personal health care diagnosis systems. The performance of classification, clustering and disease prediction are influenced by the prior stage where similarity between time series is performed. Physiologic signals vary even within the same patient, so an analysis of their possible variation without affecting future clustering accuracy is hereby addressed. Commonly employed methods of measuring similarity between time series were tested on longer data segments than the typical cardiac cycle envisaging its use integrated on personalized health care cardiovascular diagnosis systems. Euclidean distance, Discrete Wavelet Transform, Discrete Fourier Transform, Correlation Coefficient, Mahalanobis distance, Minkowski Distance, and Dynamic Time Warping Distance were compared when 20 levels of small variations in amplitude scaling and shift, time scaling and shift, baseline variance and additive Gaussian noise are forced to the tested time series. Concentrating on the performance of the similarity methods in terms of their insensibility to small data variations results demonstrate that the time domain Correlation Coefficient is the most robust method while the Discrete Wavelet Transform is the elected one between the transform-based methods tested. Selection of a similarity method to be applied should also take into account implementation issues, namely need of data reduction to avoid computational burden, and in this case transform-based methods should be elected.