A data fusion approach to biosurveillance

A significant hurdle in biosurveillance is how to detect attacks with unknown bioagents. One reason for this is that there is a high degree of variance in medical reporting in such cases, masking statistical anomalies that might otherwise be apparent. In this article, we present an application that employs a novel two-level fusion architecture designed to contend with this problem. The lower-level fusion step detects and tracks the indication of an outbreak of some sort given a set of noisy patient records, based on an information retrieval technique called latent semantic analysis. The higher level fusion step then determines the type of outbreak, based on dynamic Bayesian networks that model cause-effect interrelationships among several sources of information such as terrorist activities, environment, diseases and symptoms. We have developed and demonstrated feasibility of the approach via simulated outbreak events provided by the BioWar simulation platform.