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.

This paper describes a Bayesian modeling approach called DUDE (Detection of Unmodeled Diseases from Evidence) that can recognize outbreaks of new forms of influenza-like illness (ILI) and create clinical characterizations of them. We demonstrate its operation on data from real-world outbreaks including an outbreak of Enterovirus EV-D68. DUDE avoids the shortcomings of previous sytems by building probabilistic models of normal (baseline) ILI activity using a large set of patient findings extracted from patient-care reports using natural language processing. (This approach to ILI detection and characterization was developed in [15]). It then looks for statistically significant deviations from baseline normal activity. Thus, DUDE does not rely on just a small set of findings (as might be extracted from patients’ chief complaints). Also, it does not model temporal patterns and therefore does not assume that the present is like the past. Finally, by removing cases of known forms of ILI (such as influenza, RSV, etc.) from the input data, it can recognize new, emergent kinds of ILI.