Algorithm for statistical detection of peaks--syndromic surveillance system for the Athens 2004 Olympic Games

No generally accepted procedure exists for detecting outbreaks in syndromic time series used in the surveillance of natural epidemics or biologic attacks.

This report evaluates the usefulness for syndromic surveillance of the Pulsar approach, which is based on removing long-term trends from an observed series and identifying peaks in the residual series of surveillance data with cutoffs determined by using a combination of peak height and width.

Simulations were performed to evaluate the Pulsar method and compare it with other approaches. The daily syndromic counts in emergency departments of four major hospitals in the Athens area during August 2002-August 2003 were analyzed for two common syndromes. A standardized residual series was generated by omitting trends and noise in the original data series; this series was examined for the presence of peaks (i.e., points having magnitude higher than at least one of three probabilistically determined cutoffs). The whole process was iterated, and the baseline was recalculated by assigning reduced weight to the identified peaks.

For the specific simulation schema used, the Pulsar method fared well when compared with other approaches in meeting the performance criteria of sensitivity, specificity, and timeliness.

Although the suggested algorithm needs further validation regarding the correspondence between detected peaks and true biologic alerts, the Pulsar technique appears effective for observing peaks in time series of syndromic events. The simplicity of the algorithm, its ability to detect peaks based not only on height but also on width, and its performance in the simulated data sets make it a promising candidate for further use in syndromic surveillance.