A Bayesian approach for detecting a disease that is not being modeled

Aronis, John M.; Ferraro, Jeffrey P.; Gesteland, Per H.; Tsui, Fuchiang; Ye, Ye; Wagner, Michael M.; Cooper, Gregory F.

The simplest univariate detection algorithms [8] track a time-series of a single value, such as emergency department visits or thermometer sales, and look for significant deviations from a baseline level of expected activity. Multivariate systems [9] combine several indicators into a single compound indicator in seeking to increase performance. However, these systems suffer from two problems. First, if an outbreak of a new disease occurs during a larger outbreak of a known disease, it might not be noticed. For instance, imagine if a new disease causes fever. People with this disease might purchase thermometers. However, if an outbreak of this new disease occurs during a large outbreak of influenza the increased thermometer sales due to the new disease would be overshadowed by the number of thermometer sales due to influenza. Second, they assume that the future will be like the past, and that outbreaks of influenza and other known diseases occur at the same time each year. For instance, suppose that the system expects increased thermometer sales during the month of January because that is when outbreaks of influenza typically occur but, in fact, there is no January outbreak of influenza in the current year. Then, an increase in thermometer sales in January due to an outbreak of a new disease might well be attributed to the expected outbreak of influenza and dismissed.