A Multivariate Bayesian Method for Spatial Biosurveillance

• Activity Number: 318
• Type: Invited
• Date/Time: Tuesday, July 31, 2007 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statisticians in Defense and National Security
• Author(s): Daniel B. Neill*+ and Andrew W. Moore and Gregory F. Cooper
• Companies: Carnegie Mellon University and Carnegie Mellon University and University of Pittsburgh
• Keywords: cluster detection ; data mining ; biosurveillance

Abstract

We present a new method for spatial cluster detection, the "multivariate Bayesian scan statistic" (MBSS). MBSS combines information from multiple data streams in a Bayesian framework, computing the posterior probability of each outbreak type in each spatial region. MBSS learns a Gamma-Poisson model for each region from historical data, and models the effects of each outbreak type on each stream using expert knowledge. We evaluated MBSS on simulated outbreaks injected into three streams of OTC sales data (cough, fever, thermometers) for western PA. At 1 fp/month, MBSS detected outbreaks in 1.59 days, as compared to 2.23, 2.59, and 1.99 days for univariate Bayesian detectors monitoring each stream. MBSS was also able to characterize outbreaks by pinpointing the affected region, and accurately distinguished between different outbreak types. This work was supported by NSF grant IIS-0325581.