The purpose of infectious disease surveillance is to inform the public health policy makers on the incidence and trends of infectious diseases and to trigger appropriate actions to control infectious disease outbreaks. The enormous amount of data collected with automated laboratory-based surveillance systems require automated algorithms for detecting unexpected aberrations that may signal infectious disease outbreaks. In this paper, we explore the potential of hierarchical time series models to detect deviations from expected incidence. As these count data are extremely noisy it can be expected that these models are suitable to detect signal from noise and accommodate for possible autocorrelation. The proposed procedure consists of three steps; (1) the model parameters are estimated by empirical Bayes on a training period of, e.g. a year; (2) the expected values are updated for small time steps (e.g. daily) as new data arrive; (3) threshold levels are updated conditionally on the past expected values and an alarm is triggered when the threshold level is exceeded. To test the potential of the models we estimated sensitivity, specificity and timeliness on simulated time series and compared the results with an alternative approach of a linear regression model adjusted for trends and season. We also used two observed series for Rubella notifications and Salmonella infections and compared our findings with the expert opinions on these series. The hierarchical time series models approach shows high sensitivity and specificity and correctly identifies outbreaks at an early stage. This, in our opinion, makes the proposed model a reliable tool for adequate automated detection of infectious disease outbreaks.

Copyright 2006 John Wiley & Sons, Ltd.