Monday Research Seminar Series
January 14, Daniel B. Neill
Noon, Hamburg Hall, Room 1502
A Multivariate Bayesian Method for Spatial Biosurveillance
Automatic disease surveillance systems are essential for early detection of public health threats such as avian influenza or bioterrorism. In past work, we have developed a system which monitors multiple streams of electronically available public health data (including hospital visits and pharmacy sales) and automatically detects emerging outbreaks of disease. This system first forecasts the expected counts for each data stream for each zip code using historical data, and then detects spatial clusters where the observed counts are significantly higher than expected.
Our recently proposed "multivariate Bayesian scan statistic" (MBSS) method combines information from the multiple data streams in a Bayesian framework, computing the posterior probability of each outbreak type in each spatial region. This multivariate approach has many advantages over typical univariate detection approaches, including high detection power, fast
computation, easy interpretability, and the ability to incorporate prior knowledge of outbreak size, shape, and impacts on the different data streams. Most importantly, MBSS allows faster and more accurate detection by integrating information from the multiple streams, and enables us to model and differentiate between multiple potential causes of an outbreak.
This talk will present an overview of the MBSS method and evaluate its performance on a variety of semi-synthetic outbreak detection scenarios, using real hospital and pharmacy data from Allegheny County. I will also discuss how incremental learning (both passive and active) can be incorporated into the MBSS framework and used to improve detection performance.
