Bayesian Biosurveillance

1
(No Transcript)
2
Bayesian 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
3
Outline

• Biosurveillance goals
• Bayesian biosurveillance
• A Bayesian biosurveillance model (PANDA)
• Summary and future plans

4
Biosurveillance 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

5
Bayesian Biosurveillance
6
PANDA 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.

7
Example 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
8
Person Model
9
Some 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.

10
Equivalence Classes
The 1.4M people in the modeled population can be
partitioned into approximately 48,000
equivalence classes
11
Modeling 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.

12
Tractably 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
13
Simple 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
14
Evaluation

• 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.

15
Results
16
PANDA Spatial Model
17
Spatial Model
18
Optimized Spatial Model
19
Optimized Spatial Model
20
Optimized Spatial Model Versus a Control Chart
Method
21
Timing 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

22
Summary

• 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.

23
Future 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

24
Thank you
RODS Laboratory http/www.health.pitt.edu/rods/ B
ayesian Biosurveillance http//www.cbmi.pitt.edu/
panda/
25
(No Transcript)