Epidemiological response to syndromic surveillance signals

The epidemiological response to syndromic surveillance data must be tempered by the following considerations. It is not yet known how accurately either the syndromes themselves or the data used to define them predict or correlate with the target conditions/diseases under surveillance. In addition, because of the need for maximal sensitivity, the positive predictive value of an alarm signal for biological terrorism is by necessity going to be extremely low. It is not known what the positive predictive value of a syndromic surveillance signal is for other naturally occurring conditions of public health importance. Finally, the statistical methods used to analyze and interpret syndromic surveillance data are new and have not been sufficiently evaluated under “real world” conditions to understand their usefulness in public health decision making. Nonetheless, syndromic surveillance makes intuitive sense to many epidemiologists, who believe that, as the science of syndromic surveillance evolves and matures, the value of such systems will become apparent. In King County, Washington, we conduct syndromic surveillance using computerized electronic emergency department and primary care clinic databases in the form of International Classification of Diseases, 9th Revision (ICD-9) codes and chief complaint data. Aberrations in the data trigger an epidemiological response when we detect an alarm signal corresponding to a statistically significant increase over expected observations based on baseline data using the quality control cumulative sums (CUSUM) methods and those displayed in the Early Aberration Reporting System (EARS) of the Centers for Disease Control and Prevention. Investigations are also initiated for any report of an otherwise notifiable condition or unexplained death. The first step in investigating an alarm is confirmation of the signal. We “drill down” and examine the individual cases comprising the cluster that triggered the alarm to obtain additional demographic and geographic data. In this way, we have detected system errors that include duplication of individual case data and improper coding at the clinical site. If the signal is real, the ensuing steps are designed to increase the specificity of the signal to the greatest extent possible. We evaluate the absolute number of events leading to the signal and, when possible, the proportion of cases from the reporting institution. In systems with relatively small populations and fewer observations, signals frequently correspond to a small increase in target conditions. Data on whether the patient was admitted or discharged are available, and investigations are more likely to