Exploring Temporal Patterns in Emergency Department Triage Notes with Topic Models

Exploring Temporal Patterns in Emergency Department Triage Notes with Topic Models Simon Kocbek1,2, Karin Verspoor2,3, Wray Buntine4 1Computer Science & Info Tech, RMIT University, Australia 2Dept of Computing and Information Systems, The University of Melbourne, Australia 3Health and Biomedical Informatics Centre, The University of Melbourne, Australia 4Faculty of IT, Monash University, Australia skocbek@gmail.com, karin.verspoor@unimelb.edu.au, wray.buntine@monash.edu Abstract Topic modeling is an unsupervised ma- chine-learning task of discovering topics, the underlying thematic structure in a text corpus. Dynamic topic models are capa- ble of analysing the time evolution of topics. This paper explores the applica- tion of dynamic topic models on emer- gency department triage notes to identify particular types of disease or injury events, and to detect the temporal nature of these events. 1 Introduction Recording of a patient’s presenting complaints forms part of the standard triage procedure at most Australian hospital Emergency Depart- ments (EDs). The complaints are typically rec- orded as brief notes, which capture the reason the patient has come to the ED. These notes repre- sent the first point of contact of a patient with the hospital, and are a source of timely information about the health of the community served by the hospital. For instance, outbreaks of viruses or increased activity of spiders and snakes can be detected by monitoring ED visits. The range of reasons for patient visits to the ED is diverse, including both accidents or inju- ries and disease. Topic modeling of ED triage notes provides a strategy for identifying patterns in this diverse data, i.e., for abstracting over in- dividual patient visits to characterise trends in the health of the community. This abstraction gives a valuable snapshot of the health issues affecting the community. Given the temporal nature of many health and injury events, including seasonal variation in vi- ral load and day-of-the-week variation in events such as alcohol-related accidents, we expect that temporal patterns can be discerned in this data. In this work, we explore the application of dy- namic topic models on ED triage notes to identi- fy particular types of disease or injury events, and to detect the temporal nature of these events. This analysis provides insight into the changing health needs of the community. Our findings have potential application for public health sur- veillance applications, where emerging issues of public concern can be detected and an appropri- ate response can be planned. 2 Related work We treat each triage note is as one short docu- ment. It is known that it is very changeling for topic models to handle very short texts (Zhao et al. 2011) and various forms of tweet pooling on hashtag and/or author can be used to overcome this (Mehrotra et al 2013). For triage notes, how- ever, its not clear what discrete variable could be used to pool on. Therefore we have not used any methods to account for the short documents. Topic models have been only recently applied to analyse electronic health records data. Initial research suggests that the specific characteristics of the …

Content truncated to 500 words