Issue 12 - June 2003 : EFFECTIVE ANALYSIS FOR IDENTIFYING OUTBREAKS EARLY

________________________________________________________________________________

EFFECTIVE ANALYSIS FOR IDENTIFYING OUTBREAKS EARLY

While research regarding syndromic surveillance is still in its infancy, recent publications are examining the statistics of syndromic surveillance. In Boston, researchers Reis and Mandel used 10 years of emergency room visit data and chief complaint text data to evaluate the question, “What is the best way to identify a bioterrorist attack in its earliest stage?”

One of the goals of syndromic surveillance is to identify outbreaks (signals) in the midst of hundreds of regular emergency room visits (noise). For example, how many extra cases of diarrhea would constitute an outbreak? Previous models have compared daily syndrome rates with a historic forecast for each day. However, this type of daily analysis suffers from the unpredictable and seasonal nature of illness.

To tackle this problem, Reis et. al. developed several ‘filters’ for analyzing visits over time: a fixed increase in visits over seven days; a week-long steady increase; a week-long exponential increase; and a simple one-day increase (1). The idea behind this approach was that “you should put what happened today in the context of what happened the day before, and the day before…” (2)

The researchers analyzed all emergency room visits made to Children’s Hospital Boston between 1992 and 2002 (> 500,000 visits). Their goal was to find a filter that could detect an anthrax attack in the first two days by looking at patients presenting with respiratory-like or flu-like syndromes as well as total volume of ER cases. (3) Since no real anthrax outbreaks existed in the historical dataset, they introduced simulated outbreaks of 3, 7 and 14 days into the data, using an average of 20 extra “anthrax” cases/day. (4) In detection, they focused on the first few days of each simulated outbreak because “useful detection systems should be able to spot outbreaks within that time frame”.

Using the filtering models they developed, they were able to “detect their simulated outbreaks with a high degree of sensitivity and specificity”. (3) “The standard one-day filter was best for the first day of the outbreak but most vulnerable to false alarms. The moving average filter improved over time and was best on the seventh day. The linear filter was better in the earlier and middle stages of an outbreak. The exponential filter was strongest at detecting the earliest stages of the outbreak”. (2)

They concluded that the multi-day approach is significantly better than the standard one-day approach, as it increased sensitivity while maintaining specificity. (4) Since each filter had its own advantages, these researchers recommended an integrated approach to evaluating chief complaint data, incorporated multiple filters operating in parallel. Clearly, more research needs to be done on this subject. In the mean time, syndromic software should remain flexible and provide multiple threshold and/or filter options for detecting unusual clusters of disease.

1) http://web1.tch.harvard.edu/pressroom/2003/feb3.html

2) http://focus.hms.harvard.edu/2003/March7_2003/public_health.html

3) http://www.biomedcentral.com/1472-6947/3/2

4) http://www.pnas.org/cgi/content/full/100/4/1961

Back to the top