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 applied this approach to the assumed unmodeled outbreaks for the five years of data by selecting the 30% most atypical patients. Fig 6 compares doing so to using all patients. Note that the AMOC graph corresponding to the 30% heuristically selected patients is lower at nearly every point than the AMOC curve without heuristic patient selection. This means that given thresholds for each approach that led to a particular number of false alarms, the heuristic approach will almost always have fewer days to detection. (We also applied this heuristic using the 10%, 50%, 70%, and 90% most atypical patients and obtained similar, but less, improvement). Although this heuristic worked well with the assumed unmodeled outbreaks (influenza, RSV, parainfluenza, and hMPV) for the five years of data we used, it failed to identify the unmodeled outbreak near day 106 of 2014-2015 that was identified by DUDE when it used all patient records. Although the likelihood of the presence of an unmodeled outbreak did increase slightly it never exceeded 1 indicating that the presence of an unmodeled disease was likely.