Automated outbreak detection: a quantitative retrospective analysis

Abstract

An automated early warning system has been developed and used for detecting clusters of human infection with enteric pathogens. The method used requires no specific disease modelling, and has the potential for extension to other epidemiological applications. A compound smoothing technique is used to determine baseline ‘normal’ incidence of disease from past data, and a warning threshold for current data is produced by combining a statistically determined increment from the baseline with a fixed minimum threshold. A retrospective study of salmonella infections over 3 years has been conducted. Over this period, the automated system achieved >90% sensitivity, with a positive predictive value consistently >50%, demonstrating the effectiveness of the combination of statistical and heuristic methods for cluster detection. We suggest that quantitative measurements are of considerable utility in evaluating the performance of such systems.