Maëlle Salmon: Bayesian Outbreak Detection in the Presence of Reporting Delays

Nowadays, health institutions such as hospitals and public health authorities collect and store large amounts of information about infectious diseases. The collected case reports can be aggregated into time series of counts which are then analyzed by statistical methods in order to detect aberrations, since unexpectedly high counts can indicate an emerging outbreak. If detected early enough, an emerging outbreak may be controlled. However, inherent reporting delays of surveillance systems make the considered time series incomplete, which can be an impediment to the timely detection and thus to the containment of emerging outbreaks.

In the presented work, we synthesize the outbreak detection algorithms of Noufaily et al. (2013) and Manitz and Höhle (2013) while additionally addressing right-truncation of the disease counts caused by reporting delays. We do so by considering the resulting time series as an incomplete two-way contingency table which we model using negative binomial regression. Our approach is defined in a Bayesian setting allowing a direct inclusion of both estimation and prediction uncertainties in the derivation of whether an observed case count is to be considered an aberration. Altogether, our method aims at allowing timely aberration detection in the presence of reporting delays and hence underlines the need of statistical modelling to address complications of reporting systems.

In this talk, we shall first give some background about existing aberration detection methods and reporting delays in the German national surveillance system for infectious diseases. Then we shall present our method and the results of its evaluations on both simulated data and on the time series of Salmonella Newport cases in Germany 2002-2013. The presented work is joint work with Dirk Schumacher, Klaus Stark and Michael Höhle.