Evaluating Statistical Methods for Syndromic Surveillance

Abstract

Since the terrorist attacks on September 11, 2001, many state and local health departments around the United States have started to develop syndromic surveillance systems. Syndromic surveillance -a new concept in epidemiology -is the statistical analyses of data on individuals seeking care in emergency rooms (ER) or other health care settings with preidentified sets of symptoms thought to be related to the precursors of diseases. Making use of existing health care or other data, often already in electronic form, these systems are intended to give early warnings of bioterrorist attacks or other emerging health conditions. By focusing on symptoms rather than confirmed diagnoses, syndromic surveillance aims to detect bioevents earlier than would be possible with traditional surveillance systems. Because potential bioterrorist agents such as anthrax, plague, brucellosis, tularemia, Q-fever, glanders, smallpox, and viral hemorrhagic fevers initially exhibit symptoms ("present" 142 Michael A. Stoto et al. in medical terminology) of a flulike illness, data suggesting a sudden increase of individuals with fever, headache, muscle pain, and malaise might be the first indication of a bioterrorist attack or natural disease outbreak. Syndromic surveillance is also thought to be useful for early detection of natural disease outbreaks [Hen04].Research groups based at universities, health departments, private firms, and other organizations have proposed and are developing and promoting a variety of surveillance systems purported to meet public health needs. These include methods for analysis of data from healthcare facilities, as well as reports to health departments of unusual cases. Many of these methods involve intensive, automated statistical analysis of large amounts of data and intensive use of informatics techniques to gather data for analysis and to communicate among physicians and public health officials [WTE01]. Some of these systems go beyond health care data to include nonhealth data such as over-the-counter (OTC) pharmaceutical sales and absenteeism that might indicate people with symptoms who have not sought health care [Hen04].There are a number of technological, logistical, and legal constraints to obtaining appropriate data and effective operation of syndromic surveillance systems [Bue04]. However, even with access to the requisite data and perfect organizational coordination and cooperation, the statistical challenges in reliably and accurately detecting a bioevent are formidable. The object of these surveillance systems, of course, is to analyze a stream of data in realtime and determine whether there is an anomaly suggesting that an incident has occurred. All data streams, however, have some degree of natural variability. These include seasonal or weekly patterns, a flu season that appears at a different time each winter or perhaps not at all, differences in coding practices, sales promotions for OTC medications, and random fluctuations due to small numbers of individuals with particular symptoms. Fur...

A one‐sided MEWMA chart for health surveillance

“…Our proposal for a one-sided MEWMA chart is similar to that of Sonesson and Frisén 36 , who suggested the use of upper CUSUM statistics, and Stoto et al 29 , who modified the MCUSUM chart. In this proposed method, a 'barrier' is placed on the EWMA statistics given in Equation (6)…”

“…If recent data points are given a low enough weight (small in EWMA-based charts), a significant amount of data will be required in favor of the outbreak hypothesis, particularly if historic data have been contrary to that hypothesis. The modified MCUSUM of Stoto et al 29 and Fricker 30 and the one-sided MEWMA chart we have proposed are not as severely affected by the credit problem because decreases in rates do not hinder the ability of these charts to detect future rate increases.…”

“…Stoto et al 29 and Fricker 30 modified the Crosier MCUSUM chart so that it would only respond to increases in regional counts. They then compared the performance of this modified MCUSUM chart with that of multiple univariate CUSUM charts.…”

“…Stoto et al 29 and Fricker 30 also used only the upper CUSUM statistics as inputs to the MCUSUM chart, so that decreases in rates would not contribute to producing an alarm. They stated that there would be greater power in detecting increases in rates.…”

Biosurveillance: Detecting, Tracking, and Mitigating The Effects of Natural Disease and Bioterrorism

“…In spite of the gap, the standard SPC methods are sometimes applied to biosurveillance data with little modification (see, for example, Fricker (2007) or Stoto et al (2006)) and in some cases the methods are modified to attempt to account for the autocorrelation. For example, EARS applies variants of the Shewhart control chart (see equations (1) and (2) below) and a cumulative method (see equation (3) below).…”