Title: Bayesian Biosurveillance
1(No Transcript)
2Bayesian Biosurveillance of Disease Outbreaks
Gregory F. Cooper, Denver H. Dash,
John D. Levander, Weng-Keen Wong, William
R. Hogan, Michael M. Wagner
RODS Laboratory Center for Biomedical
Informatics University of Pittsburgh
3Outline
- Biosurveillance goals
- Bayesian biosurveillance
- A Bayesian biosurveillance model (PANDA)
- Summary and future plans
4Biosurveillance Detection Goals
- Detect an unanticipated biological disease
outbreak in the population as rapidly and as
accurately as possible - Determine the people who already have the disease
- Predict the people who are likely to get the
disease
5Bayesian Biosurveillance
6PANDA Population-wide ANomaly Detection and
Assessment
- PANDA models outbreaks using a causal Bayesian
network. - The causal Bayesian network in PANDA represents
probabilistic causal relationships that link
outbreak etiologies to available evidence, such
as emergency department (ED) visits. - The network is assessed from training data and
from knowledge of outbreak disease from the
literature.
7Example of a PANDA Bayesian Network that Models a
Disease Outbreak Due to an Airborne Release of
Anthrax
Global nodes
Person model
G
Interface nodes
P4
I
P1
P2
P3
8Person Model
9Some Current Model Details
- The probabilities in the person-network models
were estimated from U.S. Census data, from
historical ED data from Allegheny county, and
from the anthrax literature. - The population currently being modeled consists
of all 1.4M people in Allegheny County - The smallest region modeled is a Zip code, and
all Zip codes in Allegheny county are included.
10Equivalence Classes
The 1.4M people in the modeled population can be
partitioned into approximately 48,000
equivalence classes
11Modeling an Entire Population
people not seen in the ED
people seen in the ED
- Define the background population (e.g., using
census data) - As patients enter the ED, they get moved from
their background class to a patient class
corresponding to their symptoms. - After sufficient time passes, patients get moved
back into their background class, while other
patients get added.
12Tractably Modeling an Entire Population
Pre-compute the probability of observing the
entire background population, and replace all
equivalence classes with a single (binary) master
node
13Simple Adjustment Rule
As a person moves from equivalence class Ei to
class Ej, we can easily adjust the probability
table of E to reflect the change using
14Evaluation
- For testing, an outdoor anthrax release was
simulated using the anthrax cases output by the
BARD system. - The BARD-simulated cases of infected individuals
who visited the ED were overlaid onto actual
historical ED data. - Ninety-six such scenarios were generated and for
each the data stream of ED cases was given as
input to PANDA. - Each simulated hour, PANDA generated a posterior
probability of an anthrax outbreak. - We plotted time-to-detection versus the
false-positive rate of detection.
15Results
16PANDA Spatial Model
17Spatial Model
18Optimized Spatial Model
19Optimized Spatial Model
20Optimized Spatial Model Versus a Control Chart
Method
21Timing Results
- The following timing results are based on
monitoring historical ED data over - six days using PANDA running on an AMD Opteron
248 (2.19 GHz and 4 GB - RAM).
- Original Model4 to 5 seconds of machine time
- Original Model with Season, Day of Week, Time of
Day 15 seconds - Spatial Model 20 seconds
- Spatial Model with Season, Day of Week, Time of
Day 52 seconds
22Summary
- Biosurveillance can be viewed as ongoing
diagnosis of an entire population. - Causal networks provide a flexible and expressive
means of coherently modeling a population in
performing biosurveillance. - Inference on causal networks can derive the type
of posterior probabilities needed for
biosurveillance. - Initial results from a simulation study are
promising, but preliminary. - Inference can be computationally tractable when
modeling non-contagious disease outbreaks, such
as an outbreak due to the outdoor release of
anthrax spores.
23Future Work Includes
-
- Modeling contagious diseases
- Including over-the-counter (OTC) data
- Constructing realistic decision models about when
to raise an alert - Developing explanations of alerts
- Performing additional evaluations
24Thank you
RODS Laboratory http/www.health.pitt.edu/rods/ B
ayesian Biosurveillance http//www.cbmi.pitt.edu/
panda/
25(No Transcript)