Statistical Methods for Detecting the Onset of Influenza Outbreaks

Abstract

This paper reviews different approaches for determining the epidemic period from influenza surveillance data. In the first approach, the process of differenced incidence rates is modelled either with a first-order autoregressive process or with a Gaussian white noise process depending on whether the system is in an epidemic or a nonepidemic phase. The second approach allows us to directly model the process of the observed cases via a Bayesian hierarchical Poisson model with Gaussian incidence rates whose parameters are modelled differently, depending on the epidemic phase of the system. In both cases transitions between both phases are modelled with a hidden Markov switching model over the epidemic state. Bayesian inference is carried out and both models provide the probability of being in epidemic state at any given moment. A comparison of both methodologies with previous approaches in terms of sensitivity, specificity and timeliness is also performed. Finally, we also review a web-based client application which implements the first methodology for obtaining the posterior probability of being in an epidemic phase.

Keywords

• autoregressive modeling
• Bayesian inference
• influenza
• hidden Markov models
• public health
• temporal surveillance