Transcript Decision Theoretic Analysis of Improving Epidemic Detection

Decision Theoretic Analysis of
Improving Epidemic Detection
Izadi, M.
Buckeridge, D.
AMIA 2007,Symposium Proceedings 2007
Overview

Objective: Improve the accuracy of current
detection methods

Observation: Quantifying the potential costs
and effects of intervention can be used to
optimize the alarm function

Method: Use Partially Observable Markov
Decision Processes on the outbreak detection
method
The Surveillance Cycle*
1. Identifying
individual cases
2. Detecting population
patterns
Individual Event
Definitions
Event
Detection
Algorithm
3. Conveying information for
action
Population Pattern
Definitions
Event
Reports
Data Describing
Population
(Buckeridge DL & Cadieux G, 2007)
Intervention
Guidelines
Pattern
Report
Pattern
Detection
Algorithm
Population Under
Surveillance
Intervention
Decision
Public
Health
Action
Usual Detection Methods*

Methods are non-specific – they look for
anything unusual in the data

Design a baseline.

Define an aberration when some
statistics are more than expected values
by the baseline.
Detection Method Example*

Define a threshold for the number of Emergency
Department visits per day.

Signal an alarm when the number of ED visits per day
exceeds the threshold.
Num ber of ED Visits per Day
Number of ED Visits
50
40
30
20
10
0
1
10 19
28 37 46
55 64
Day Num ber
73 82 91 100
Sensitivity and Specificity Tradeoff
Sensitivity is the probability of alarm
given an outbreak P(A+|O+)
 Specificity is the probability of no alarm
given no out break P(A-|O-)
 Timeliness is time between outbreak and
detection
 Challenge: Increasing sensitivity and
improving timeliness decreases specificity

Approach Overview*
 Instead
of trying to improve the
detection method, ‘post-process’
the signals:
 Use
a standard detection method to
provide signals
 Feed this signal to a decision support
model to find the optimal action
Quick Introduction to POMDPs*







What goes on:
st-1
What we see:
ot-1
What we infer:
bt-1
States: sS
Actions: aA
Observations: oO
Transition probabilities: Pr(s’|s,a)
Observation probabilities: Pr(o|s,a)
Rewards: R(s,a)
Belief state: b(s)
st
at-1
ot
bt
at
Model Components

States: - True epidemic state





No Outbreak
Day1
...
Day4
Observations: Output from the detection
algorithm:


Alarm
No-Alarm
Model Components (continued)

Actions
1.
2.
3.
4.

Do nothing
More Systematic Studies (e.g. get more patient
files from ED)
More Investigation (done by human expert)
Declare outbreak
Transition and Observation Probabilities
 Calculated based on expert knowledge
Model Components (continued)*

Costs
Investigation (false and true positive)
 Intervention (false and true positive)
 Outbreak by day (false negative)

(# deaths* future earnings) + (# hospitalized *
cost of hospitalization) + (# outpatient visits * cost
of visit)
Model Components (continued)*

Rewards

Preventable loss at each day
Outbreak Detection as a POMDP*
No Outbreak
D1
D2
D3
D4
Outbreak
detected
Do nothing
Review records
Investigate cases
Declare outbreak
Experimental Design

Compare a detection method (moving
average) with and without addition of
POMDP

Consider a fixed Specificity of 0.97

The comparison is over 10 years
simulation

Not exactly clear how the data is generated
Experimental Results*
Small size outbreak
Day of Outbreak
Experimental Results*
Larger size outbreak
Day of Outbreak
Conclusion

POMDPs can improve the accuracy of the
current outbreak detection methods

We can use the potential costs and effects of
intervention to learn a decision process

P(A-|O-) = 0.97  P(A|O-) = 0.03


In every 100 days , we will have 3 false alarms!
Is this acceptable?