Learning similarities between irregularly sampled short multivariate time series from EHRs

Abstract

A large fraction of the Electronic Health Records consists of clinical multivariate time series. Building models for extracting information from these is important for improving the understanding of diseases, patient care and treatment. Such time series are oftentimes particularly challenging since they are characterized by multiple, possibly dependent variables, length variability and irregular samples. To deal with these issues when such data are processed we propose a probabilistic approach for learning pairwise similarities between the time series. These similarities constitute a kernel matrix that can be used for many different purposes. In this work it is used for clustering and data characterization. We consider two different multivariate time series datasets, one of them consisting of physiological measurements from the Department of Gastrointestinal Surgery at The University Hospital of North Norway and we show the proposed method’s robustness and ability of dealing with missing data. Finally we give a clinical interpretation of the clustering results.