This paper introduces a method of surveillance using deviations from probabilistic forecasts. Realised observations are compared with probabilistic forecasts, and the "deviation" metric is based on low probability events. If an alert is declared, the algorithm continues to monitor until an all-clear is announced. Specifically, this article addresses the problem of syndromic surveillance for influenza (flu) with the intention of detecting outbreaks, due to new strains of viruses, over and above the normal seasonal pattern. The syndrome is hospital admissions for flu-like illness, and hence, the data are low counts. In accordance with the count properties of the observations, an integer-valued autoregressive process is used to model flu occurrences. Monte Carlo evidence suggests the method works well in stylised but somewhat realistic situations. An application to real flu data indicates that the ideas may have promise. The model estimated on a short run of training data did not declare false alarms when used with new observations deemed in control, ex post. The model easily detected the 2009 H1N1 outbreak. Copyright © 2016 John Wiley & Sons, Ltd.
Keywords: Markov process; early event detection; integer autoregressive model; probability forecasts; real-time surveillance.
Copyright © 2016 John Wiley & Sons, Ltd.