There are several possible extensions to PANDA that are straightforward to implement, including (1) increasing the number of days being modeled, (2) modeling on an hourly basis, rather than a daily one, and (3) adding nodes to the model that represent prevailing wind direction and wind speed.
A more fundamental extension involves modeling a set of variables that represent the amount of over-the-counter (OTC) medication sales of a particular type (e.g., cough medication sales) per subregion (e.g., zip code) per day. Preliminary work indicates that it is feasible to develop a causal Bayesian network model that incorporates both ED data and OTC data (Wong et al., 2005).
Within the current framework, additional non-contagious outbreak diseases can be modeled, as well as non-outbreak diseases that might be easily confused with outbreak diseases. A more ambitious goal is to causally model contagious diseases, where there is much less independence among the individuals being modeled than in noncontagious diseases. Beyond the modeling issues, which are substantial, developing inference algorithms that are fast enough to permit real-time biosurveillance of contagious diseases will also be challenging. Approximate inference algorithms may be required.
Finally, much additional testing is needed of the run-time and detection performance of Bayesian biosurveillance methods, including their relative performance to other detection methods.
Was this article helpful?