Early, reliable detection of outbreaks of disease, whether natural (e.g., West Nile virus and SARS) or bioterrorist induced (e.g., anthrax and smallpox), is a critically important problem today. We need to detect outbreaks as early as possible in order to provide the best response and treatment, as well as improve the chances of identifying the source.
Outbreaks often present signals that are weak and noisy early in the event. If we hope to achieve rapid and reliable detection, it likely will be necessary to integrate multiple weak signals that together provide a relatively stronger signal of an outbreak. Combining spatial and temporal data is an important instance of such integration. For example, even though the number of patients in a given city with fever, who were seen in emergency departments (EDs) in the past 24 hours, may not be noticeably higher than average, nonetheless, for the past 12 hours, it may be significantly higher for a given neighborhood of the city.
Because of the noise in signals early in the event, early detection is almost always detection under uncertainty. In the research reported here, we use probability as a measure of uncertainty. A well-organized probabilistic approach allows for the rational combination of multiple, small indicators into a big picture. Since the interrelationships among risk factors, diseases, and symptoms often are causal, a causal representation provides a natural approach to modeling. This chapter concentrates on causal modeling of how the data could be produced by various hypothesized outbreak diseases. The more we know about the causal relationships among the risk factors, the outbreak disease etiologies, and the clinical presentations of those outbreak diseases, the more compelling is the causal modeling approach.
This chapter focuses on the use of causal Bayesian networks to represent causal relationships under uncertainty. The chapter starts with a brief overview of Bayesian modeling in general and then causal Bayesian networks in particular. In order to illustrate Bayesian modeling using Bayesian networks, the remainder of the chapter describes a detailed example in which a causal Bayesian network is used to model an entire population of people (Cooper et al., 2004).1
This section provides a brief background on Bayesian modeling and inference (Bernardo and Smith, 1994) and on Bayesian networks (Neapolitan, 2004).
Was this article helpful?