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 WSARE (What’s Strange About Recent Events) system [10] addresses these issues by representing the joint distribution of patient data with a Bayesian network that includes several environmental attributes to represent influenza activity, season, weather, etc. and several response attributes to represent patient attributes such as age, gender, location, and reported symptom (which is similar to a patient’s chief complaint and takes a value from none, respiratory problems, nausea, or rash). The network is conditioned on the current values of the enviromental attributes to create a conditional joint distribution of response variables for the current day. Thus, the conditional joint distribution represents what would be expected for the current day if there are no outbreaks of new diseases. Thus, WSARE does not take a strictly technical approach to predicting cyclic patterns, but rather incorporates current conditions. The WSARE system then searches for rules that describe significant differences between the actual current data and the conditional joint distribution. For instance, if there are many patients with fever in the current data, but the conditional joint distribution predicted few patients with fever, WSARE would report this. The WSARE system has two important shortcomings, however. First, as with time-series methods, if an outbreak of a new disease occurs during a large outbreak of influenza it might not be noticed. Second, it can be misled by outbreaks of other known diseases such as RSV, parainfluenza, hMPV, etc. For instance, if the environmental attribute influenza activity is low, but there is an outbreak of RSV, the resulting surge of patients with respiratory ailments would cause a false alarm. (We note that these shortcomings could be corrected by adding additional enviromental attributes for each known disease and additional response attributes for a set of clinical findings sufficient to distinguish between different respiratory illnesses).