2014.ISDS.Abstracts.Final.pdf


ISDS Annual Conference Proceedings 2014. This is an Open Access article distributed under the terms of the Creative Commons Attribution-
Noncommercial 3.0 Unported License (http://creativecommons.org/licenses/by-nc/3.0/), permitting all non-commercial use, distribution, 
and reproduction in any medium, provided the original work is properly cited.

ISDS 2014 Conference Abstracts

A Term-based Approach to Asyndromic Determination 
of Significant Case Clusters

Howard Burkom*, Yevgeniy Elbert, Christine Piatko and Clay Fink

Johns Hopkins Applied Physics Laboratory, Laurel, MD, USA

Objective
Explain and demonstrate the performance of a statistical method 

for detection of anomalous terms in pooled, contiguous blocks of free-
text chief complaints from a health facility with emergent or urgent 
care capability.

Introduction
Biosurveillance systems commonly depend on free-text chief 

complaints (CC)s for timely situational awareness. However, 
diagnosis codes may not be available soon enough and may have 
uncertain value because they are assigned for billing purposes rather 
than for population monitoring. Existing systems use syndrome 
categories to classify records based on these free-text fields. A 
syndromic cluster determination method (TOA) based on patient 
arrival times has been implemented in versions of ESSENCE and in 
NCDETECT [1]. While effective for finding case clusters whose CC 
terms are classifiable into syndromes, TOA implementations do not 
find clusters whose CC terms share only uncategorized terms.

Natural language processing literature is rich with topic-detection 
methods, many involving medical ontologies and language models. 
However, identifying clusters of CCs poses several challenges. Free-
text CCs often contain only a few words, include abbreviations that 
vary among institutions, and evolve with usage and presentation 
trends. Our approach pools CCs into contiguous time blocks and uses 
a statistical hypothesis test to seek current terms that are anomalous 
relative to their occurrence in a large sliding baseline. Sets of 
anomalous terms (if any) are then presented for further investigation.

Methods
Our development used a collection of emergency department (ED) 

patient records with CCs spanning 7 years from 15 hospitals in the 
National Capital Region. For pre-processing, spelling correction 
was applied followed by removal of English stop-words and of the 
common words “pain”, “left”, and “right”. For anomalous word 
detection, we applied Fisher’s exact test to the count of each word in 
the tested 8hour block of CCs, based on a 30-day sliding baseline of 
pooled CCs ending exactly 7 days before the current test block. Using 
Fisher’s exact test, anomalous words from the current 8-hr block may 
be presented to human health monitors up to 3 times each day.

We tested this method to detect two anomaly types: a small number 
of instances of rare words with little or no representation in the sliding 
baseline, and disproportionately large numbers of common words. 
Multiple alerting thresholds, block sizes, and hypothesis tests were 
tested to achieve both types of detection with at most a handful of 
anomalous words expected from each block. For scenario-based 
detection, sets of records were injected using stochastic draws from 
distributions of demographic fields, of words associated with chosen 
outbreak types, and of date/time distributions covering from 6 hours 
to 3 days and kept consistent with diurnal ED visit patterns.

Results
The methods were tested on CCs from individual hospitals and on 

pooled data from 15-20 hospitals. Individual anomalous word tests 
were conducted to test the ability to detect excess clusters of rare, 

moderately common, and very common terms, respectively. These 
terms were injected into each 8-hour block over 7 years using sliding 
baselines with no injects.

As an alert burden measure for 15 single hospitals, the method 
returned no anomalous words in 60-90% (depending on the facility) 
of the tested (>7,600 consecutive) 8-hr blocks for a pvalue threshold 
of 0.01. In 95-100% of these blocks, fewer than 5 anomalous words 
were found. Multiple ROC-like curves showed practical sensitivity. 
For example, using data from the busiest single ED tested, 3 additional 
occurrences of “exposure” were detected in >90% of the tested blocks 
with less than 2 anomalous words identified per block.

Conclusions
The presented method is a practical, understandable way to monitor 

single care facilities for CC clusters of concern based on unusually 
high occurrence of rare or common terms that need not be related 
to syndromes. Routine implementation requires a human monitor 
to inspect CCs containing the anomalous terms and make follow-up 
decisions. A prototyped combined visualization shows recent blocks 
of these anomalous CC terms with syndromic time-of-arrival alerts 
and unusual groupings of demographic strata.

Keywords
chief complaint; Fisher’s Exact Test; asyndromic; cluster

References

Deyneka L, Xu Z., Burkom H., Hicks P., Benoit S., Vaughan-Batten H., 
Ising A. Finding time-of-arrival clusters of exposure-related visits 
to emergency departments in contiguous hospital groups. Emerging 
Health Threats Journal. 2011; 4:11702. doi: 10.3402/ehtj.v4i0.11702.

*Howard Burkom

E-mail: howard.burkom@jhuapl.edu    

Online Journal of Public Health Informatics * ISSN 1947-2579 * http://ojphi.org * 7(1):e11, 2015