When biosurveillance monitoring is begun, the set of evidence e represents background information about the population. Currently, as background information, we use U.S. Census data to provide the age, gender, and home zip code information for the people in the region being monitored for disease outbreaks.

After P(e 11) is calculated once for the entire population, we can apply Eq. 5 and update this quantity incrementally as we observe people enter the ED in real time. As we observe a person from equivalence class Qk enter the ED, we find the class Qij that this person must have originated from in the background population. For example, if we observe a patient in the ED with the following attributes: Qk = {Home Zip=15260, Age=20-30, Gender=F, Date Admitted=today, Respiratory symptoms =yes}, then we know that she originated from the background class Qj = {Home Zip=15260, Age=20-30, Gender=F, Date Admitted=never, Respiratory symptoms= unknown}.

Applying the incremental updating rule allows us to reduce the number of updates that need to be processed each hour to dozens (= rate of patient visits to all the EDs in the region) rather than the millions (= the number of people in the regional population).

By caching equivalence classes and applying incremental updating, we can process an hour's worth of ED patient cases (about 26 cases) from a region of 1.4 million people in only 11 seconds using a standard Pentium III PC and the Hugin BN inference engine v6.2 (Hugin, 2004). Thus, there is enough computing reserve to "keep ahead'' of the real-time data, even when in the future we extend our model to be considerably richer in detail, and we widen the geographic region being monitored for a disease outbreak.

Was this article helpful?

0 0

Post a comment