Abstract
ObjectiveThe objective of this presentation is to describe the new word alertcapability in ESSENCE and how it has been used by the FloridaDepartment of Health (FDOH). Specifically, this presentation willdescribe how the word alert feature works to find individual chiefcomplaint terms that are occurring at an abnormal rate. It willthen provide usage statistics and first-person accounts of how thealerts have impacted public health practice for the users. Finally,the presentation will offer future enhancement possibilities and asummary of the benefits and shortcomings of this new feature.IntroductionSyndromic surveillance systems have historically focused onaggregating data into syndromes for analysis and visualization. Thesesyndromes provide users a way to quickly filter large amounts ofdata into a manageable number of streams to analyze. Additionally,ESSENCE users have the ability to build their own case definitionsto look for records matching particular sets of criteria. Those user-defined queries can be stored and analyzed automatically, along withthe pre-defined syndromes. Aside from these predefined and user-defined syndromic categories, ESSENCE did not previously providealerts based on individual words in the chief complaint text that hadnot been specified a priori. Thus, an interesting cluster of recordslinked only by non-syndromic keywords would likely not be broughtto a user’s attention.MethodsIn the FDOH ESSENCE system a new detection feature wasdeveloped to trigger alerts based on anomalous occurrence of termsin chief complaints.1This feature used Fisher’s Exact Test to testfrequencies of individual chief complaint terms relative to all termsin a 1-month baseline. The feature used a 7-day guard-band, andautomatically switched to an efficient chi-square test for sufficientlylarge term counts. A term triggered an alert if its p-value≤10E-4.This algorithm was then run on chief complaint sets both by hospitaland by region, with region assignment according to patient zip code.Results were then displayed in new visualizations showing alerts inword cloud and line listing form. Additionally, users were given theoption to ignore stop words, syndromic terms, and a user-created listof ignorable words in order to focus on words of greater interest.ResultsThe result of using the tool since June 2016 has seen three majorbenefits. First, the original intent for the system to notify users ofabnormal word clusters has proven useful. Users have been able to seeterms such asDisaster, ShelterandFireworkswhich were not part ofany prior syndromes and use these notifications to investigate possibleissues. The second benefit found by users was the ability to find newmisspellings or abbreviations commonly used by hospitals. The termsZykaandGLF(Ground Level Fall) are examples of these. Finally,the system has helped discover new trends in hospital processes. Forexample, the tool has helped discover first person and non-Englishphrases in the chief complaint. This observation led to the discoverythat some hospitals are using kiosks or mobile phone apps to allowpatients to enter their own chief complaints.ConclusionsThe word alert feature has provided value to the users of FDOHESSENCE. While accomplishing its initial goal of triggeringabnormal non-syndromic term usage, the additional ability to findnew misspellings and abbreviations may have even larger impact bykeeping syndrome and subsyndrome definitions up-to-date over timefor traditional syndromic alerting. Beyond these current benefits,additional visualization enhancements are under consideration.Additionally, the resources required to perform the detection aresubstantial, and implementation improvements are under developmentto improve the performance and enable more advanced free-textanomaly detection.