A hybrid hierarchical Bayesian model for spatiotemporal surveillance data

Abstract

Due to the low signal-to-noise ratio and high-dimensional structure, spatiotemporal data analysis is challenging. In outbreak detection, the assumptions for control charts, including independence, normality, and stationarity, are often violated in syndromic surveillance data. We develop a novel hybrid hierarchical Bayesian model through the combination of the Dirichlet process and particle filters to resolve these issues. We use a modified adjacency matrix as the observation matrix in the Markovian state-space model. This methodology achieves dimension reduction and computational efficiency and enjoys a superior detection performance with the ability to incorporate online updating for data streaming applications. Our data set is derived from the Indiana Public Health Emergency Surveillance System. It consists of surveillance data on emergency room visits for influenza-like and respiratory illness from 2008 to 2010. We are able to detect the 2009 H1N1 outbreak as well as the seasonal influenza outbreaks. Numerical results show that our Dirichlet process/particle filter models improve the outbreak detection performance in both simulation and real data analysis.