PANDA  ("Population-wide ANormaly Detection and Assessment")

PANDA is a spatio-temporal detection algorithm designed to model and detect disease outbreaks, such as an airborne release of anthrax. A unique feature of PANDA is its use of a causal Bayesian network that models each individual in the population of a given region being monitored. In contrast, most existing detection algorithms take a population-based approach by operating on data aggregated over the entire population. Modeling at the individual level has several advantages. First, it allows different data about individuals to be incorporated into the analysis of an outbreak. If we know more information about one person than another, we can represent that difference. Second, it allows us to model what we know (probabilistically) about the likely patterns of presentation of known diseases, such as airborne anthrax. Third, it facilitates the fusion of different data sources, because such data originate from the individuals in the population that are being explicitly modeled.

One of the key difficulties with PANDA’s individual-based modeling approach, however, is the high computational expense. Since each individual in the population is modeled explicitly, the overall model may contain millions of variables. PANDA addresses this computational problem through the use of several optimization and modeling techniques. It presently is able to monitor in real time a region containing over one million people.

The current version of PANDA is being tested as a research prototype. It takes Emergency Department (ED) data as input, which includes patient age, gender, the date the patient was seen at the ED, the patient home zip code, and whether or not the patient had a respiratory chief complaint; it then outputs as a function of time a probability of an anthrax disease outbreak, as well as the likely location of the hypothesized release.

Ongoing development of PANDA includes (1) combining over-the-counter medication sales data with ED data in order to perform outbreak detection of an outdoor airborne release of anthrax spores and (2) incorporating a more detailed spatio-temporal model of anthrax infectivity to use in outbreak detection.

References

Cooper, G.F., Dash, D. H., Levander, J. D., Wong, W-K, Hogan, W. R., and Wagner, M. M. Bayesian Biosurveillance of Disease Outbreaks. Proceedings of the Conference on Uncertainty in Artificial Intelligence, 2004.