Analyzing%20over-the-counter%20medication%20purchases%20for%20early%20detection%20of%20epidemics%20and%20bio-terrorism

1
Analyzing over-the-counter medication purchases
for early detection of epidemics and bio-terrorism

• by Anna Goldenberg
• Advisor Rich Caruana
• Note Sponsored by CDC Grant

2
Problem Statement

• Long history of epidemics and bio-terrorism
  attacks
• no good early detection system!

3
Existing Solutions

• Enforced by Department of Health
• Quarantine there has to be enough evidence of
  mass sickness
• Sanitation always helps but what if its an
  intentional release of bioagent?
• Immunity
• Vaccination
• Computer Surveillance Systems

- do not prevent from new strains
- do not prevent from new strains
4
Existing Solutions

• Enforced by Department of Health
• Quarantine there has to be enough evidence of
  mass sickness
• Sanitation always helps but what if its an
  intentional release of bioagent?
• Immunity
• Vaccination
• Computer Surveillance Systems
• System for clinicians to report suspicious trends
  of possible bio- terrorist events
• assessing the current capacity of hospitals and
  health systems to respond to a bio-terrorist
  attack
• evaluating and improving linkages between the
  medical care, public health, and emergency
  preparedness systems to improve detection of and
  response to a bio-terrorist event

- do not prevent from new strains
5
Gap

• Fault
• Existing CBSS rely on medical records
• may not be early enough! (anthrax)

6
Gap

• Fault
• Existing CBSS rely on medical records
• may not be early enough! (anthrax)
• Solution
• Create a system based on non-specific
• syndrome data, for e.g. over-the-counter
  medications

7
Proposed Framework
Data Preprocessing
Merge to get final prediction
Smoothed Model
Decomposition
Prediction of each component
Real-time data gt threshold
NO
YES
WARNING! POSSIBLE BEGINNING OF AN EPIDEMIC
8
Proposed Framework
Data Preprocessing
Merge to get final prediction
Smoothed Model
Decomposition
Prediction of each component
Real-time data gt threshold
NO
YES
WARNING! POSSIBLE BEGINNING OF AN EPIDEMIC
9
Smoothed Model

• Smooth original data by using DCT
• and removing small coefficients
• that correspond to noise

DCT
rms0.0798
k1,..,N, N length of data vector
TOO SMOOTH!
rms 0.1055
10
Proposed Framework
Data Preprocessing
Merge to get final prediction
Smoothed Model
Decomposition
Prediction of each component
Real-time data gt threshold
NO
YES
WARNING! POSSIBLE BEGINNING OF AN EPIDEMIC
11
Decomposition using wavelets
12
Proposed Framework
Data Preprocessing
Merge to get final prediction
Smoothed Model
Decomposition
Prediction of each component
Real-time data gt threshold
NO
YES
WARNING! POSSIBLE BEGINNING OF AN EPIDEMIC
13
Predictions
Since each component is smooth using linear
methods, such as AR, for predictions of each
component
14
Proposed Framework
Data Preprocessing
Merge to get final prediction
Smoothed Model
Decomposition
Prediction of each component
Real-time data gt threshold
NO
YES
WARNING! POSSIBLE BEGINNING OF AN EPIDEMIC
15
Comparison step
Data falls under the threshold -gt declare normal
flow. No flag is raised. Note in reality no
outbreak at that time
16
Proposed Framework
Data Preprocessing
Merge to get final prediction
Smoothed Model
Decomposition
Prediction of each component
Real-time data gt threshold
NO
17
Why so many steps?

• Smoothing
• original data is too hard to predict
• little confidence in prediction
• Decomposition
• even after smoothing too complicated for
  regular TSA tools to predict
• Main Reason need as much confidence in
  our model as possible
• lives may depend on this!

18
Results

• Ran the system according to the framework with
  different thresholds (as in the legend)

Detected strong epidemic 8 days early, weak one
2 days early had one false alarm with
threshold set as 4 above prediction
19
Complications

• Hard to make predictions around big holidays. It
  is possible that people stock up at that time
• Lack of detailed data concerning real outbreaks
• Difficulty in distinguishing between very early
  prediction and false alarms
• So far, need to consult an expert on the issues
  above.

20
Future Work

• Analyze the lower bound on accuracy of the
  prediction
• Incorporate expert knowledge into the process,
  for e.g. remove known periodicities
• Predict based on a selection of products, not
  just one category
• Set threshold to be the function of cost when
  acted upon a false alarm

21
Questions?