Time-series as background data for relating medical diagnoses terms

Abstract

Relating terms from different ontologies or identifying the most relevant entry in an ontology for a given term, is an important task in various settings involving the use of ontologies. Often the task of relating terms is achieved by considering the instance level matchings within the ontologies being aligned, or using an external common ontology for indirect linking, or using annotated text corpus. In this paper, we focus on a variant of this problem that occurs in relating medical diagnosis terms. We propose a novel unsupervised approach that exploits the availability of time-series data medical events of patients during their stay in intensive care unit (ICU). Our method, called Decree, is evaluated using a large-scale real-world medical repository of ICU data including event data from laboratory test measurements to quantify the relationship strength between terms from a given ontology. We further outline how Decree can be used to assign diagnoses terms in case of unlabeled pathology as well. We show that Decree can discover better quality relationships and is more scalable than state-of-the-art time-series techniques.