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.

We create two windows on the data. The baseline window includes patient records from a controlled time period when no outbreaks of new viruses are suspected. DUDE collects statistics about rates of symptoms among patients in the baseline window to model background ILI. The monitor window includes patient records from a later time period. DUDE then computes the likelihood of the data in the monitor period assuming that only background ILI is present, and the likelihood of the data in the monitor period assuming that both background ILI and a different unmodeled ILI is present. We then compute the odds that an unmodeled ILI is present. Note that our data are collected in hospitals. The number of patients with a disease who go to a hospital is certainly related to the number of cases in the general population. Even if this is a simple linear relationship it is hidden since we have no way to directly count the number of people in the general population with a disease We avoid the difficulty of determining this relationship by working solely with ED data.