Typical approaches to monitoring ED data classify cases into pre-defined syndromes and then monitor syndrome counts for anomalies. However, syndromes cannot be created to identify every possible cluster of cases of relevance to public health. To address this limitation, NC DETECT’s approach clusters cases by arrival times and monitors the textual chief complaint data associated with each identified cluster for relevant similarities [1]. This approach is time consuming and limited in its ability to detect emerging outbreaks that are dispersed across time. A new method is needed to automatically identify clusters of interest that would not be detected by existing syndromes. Clusters may be based on symptoms, events, place names, arrival time, or hospital location. The NC DPH dataset describes 198,511 de-identified ED visits over one year at 3 North Carolina hospitals. The data include chief complaint, altered date and time of arrival, hospital A/B/C, and age group. About 40 simulated outbreaks were injected into the data set by the NC DETECT team. For example, an inject cluster might consist of 4 patients who report getting sick after eating at a particular restaurant.
Objective
We apply a novel semantic scan statistic approach to solve a problem posed by the NC DETECT team, North Carolina Division of Public Health (NC DPH) and UNC Department of Emergency Medicine Carolina Center for Health Informatics, and facilitated by the ISDS Technical Conventions Committee. This use case identifies a need for methodology that detects emerging, potentially novel outbreaks in free-text emergency department (ED) chief complaint data.