TUTORIAL T6

1
TUTORIAL T6 Theory and Practice of Outbreak
Detection
AMIA Annual MeetingSaturday, November 8, 2003
800 - 430 pm
2
The Instructors
Andrew Moore
Mike Wagner
John Loonsk Bill Hogan
Marc Overhage
3
Schedule
I. BACKGROUND (Wagner) 30 minutes II. DATA
(Wagner) 1 hour -BREAK- III. ALGORITHMS
(Moore) 1.5 hours (spans the lunch break)
-LUNCH- IIIb. ALGORITHMS CONTINUED
(Hogan) .5 hours IV IMPLEMENTATIONS (Overhage)
-BREAK- FUTURISTIC PANEL V. CDC
IMPLEMENTATIONS (Loonsk) VI Wrap up
4
I. OVERVIEW
5
THINGS HAVE REALLY CHANGED Defend the borders by
checking everyones temperature
August 24, 2003 Taipei
6
Defend the hospital by checking everyones
temperature and symptoms at the front door
August 24, 2003 Taipei
7
Medical Computer Scientist (aka Medical
Informatician)
Medical Informatics System Being Demod
Computer Scientist
February 5, 2002
8
Public Health Law Wisconsin and Utah

• Wisconsin
• Act 109 a pharmacist or pharmacy shall report
• An unusual increase in the number of
  prescriptions dispensed or nonprescription drug
  products sold for the treatment of medical
  conditions specified by DHFS by rule.
• An unusual increase in the number of
  prescriptions dispensed that are antibiotic
  drugs.
• The dispensing of a prescription for the
  treatment of a disease that is relatively
  uncommon or may be associated with bioterrorism

• Utah
• Act 26-23b-105. A pharmacist shall report
• an unusual increase in the number of
  prescriptions filled for antimicrobials
• any prescription that treats a disease that has
  bioterrorism potential if that prescription is
  unusual or in excess of the expected frequency
  and
• an unusual increase in the number of requests for
  information about or sales of over-the-counter
  pharmaceuticals

Michigan too!
9
Scope

• Medical informatics is 90 sociology 10
  technology (Al Pryor, LDS Hospital HELP system
  pioneer)
• Real-time outbreak detection is also 90
  sociology 10 technology

Good news we can adequately cover the 10
technical part in 6 hours!
I will give 5 minute on the other 90 at the very
end!
10
I am Drawing from Several Reports

• Nations Current Capacity (Dato, Wagner et al.
  2001 88 pages.
• Representative Threats for Research in Outbreak
  Detection (Wagner, Dato et al. 2003)
• Emerging Science of Very Early Detection of
  Disease Outbreaks (Wagner et al. 2001)
• How Outbreaks are Detected in US (Dato, Wagner,
  Fahoudar 2003)
• (see www.health.pitt.edu/rods/publications.htm )

11
Very Selective, Myopic, and Brief History of
Public Health Surveillance

• Public health practice dates to the Greeks
• 1800s John Snow used mapping of cases to identify
  cause of cholera outbreak in London
• 1950 Digital computer invented.
• 1999 concerns about emerging diseases and
  bioterrorism sparked research and system building
  designed to exploit potential of information
  technology

12
Basic Concepts and Working Definitions

• Outbreak anomalous density of cases
• Detection Noticing the existence of an anomaly
• E.g., one case of very unusual illness
• E.g., more sick people than usual
• Detection Characterizing it
• Which pathogen
• Source
• Scope (people, animals)
• Environmental scope
• Route of transmission
• Response E.g. quarantine, treatment,

13
There are a lot of organisms. Which one to start
doing research with?
14
Our Starting Point--Anthrax Scenario

• WHAT IF
• 100,000 people are exposed
• Onset of illnesses occurs over days 1-7
• The costs are
• Treatment of sick
• Prophylaxis of healthy (exposed and unexposed)
• Future earnings lost through deaths, valued at
  approximately 790,000 per
• Mass treatment occurs on days 0, 1, 2, 3, 4, 5, 6
  or 7 and it has 90 efficacy

Worst case
Best case
Kaufmann, The economic impact of a bioterrorist
attack Are prevention and postattack
intervention programs justifiable EID 3(2)83-94,
1997.
15
If Public Health Surveillance Were Implemented
Like Missile Defense
Release
TIME
5 days
Teich, Wagner et al, JAMIA 9(1), 2002
16
What Might Instrumenting the Community Look Like?
Clinical
Pre Clinical Behavior
Animal
PBMs
OTC Electrolytes
Vets
Orders for tests
OTC Eqpt
Zoos
Poison Cntrs
OTC Meds
Animal Cntl
911 Calls
Absenteeism
Agribusiness
Utility use
EMS Runs
Phone traffic
ED Visit Chief Complaints
Web Queries
Radiograph reports
Microbiology
Calls to triage center
Sentinel MD
Diagnoses
Cultures
Coroner
Health dpt. ecords
17
If Missile Defense Were Implemented Like Public
Health Surveillance
Missile Launch
Missile Impact
TIME
25 Minutes
Astute clinician B notices and fills out disease
reporting form the next day
18
Anthrax Approach
What we do now for Anthrax
Chief Complaint
Chief Complaint
Chief Complaint
Chief Complaint
Chief Complaint
Chief Complaint
Chief Complaint
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
Spatial and temporal analysis to detect
overdensity of cases in a zip code or larger
region
19
Detection-Response Coupling
DETECTION
Release
P0.0001
P.05
P0.01

RESPONSE
Mobilize NPS
Mobilize Pittsburgh Pharmaceutical Stockpile
Ask Poison Center to collect data
20
Biosurveillance 2005 Maximally accelerated
detection-response cycle
First Hint of Trouble -statistical analysis of
data -astute observer -definitive diagnosis of
new or terrorism organism
21
How to Deal with a very Big Problem Space?
22
Which Organisms?
Contagious person-to-person aerosol rubella,
zoster Premonitory release Boca
Anthrax Foodborne Salmonella sp. Waterborne
Cryptosporidium VectorborneWest Nile, Malaria,
Lyme Continuous release of bioaerosol
Legionella, pollen Building contamination
CO Sexually transmitted HIV, N. gonorrhea, Hep
C Large scale bioaerosol NO OUTBREAKS Dato,
Wagner, et al. 2001      

• The Nation's Current Capacity for the Early
  Detection of Public Health Threats including
  Bioterrorism. June 8, 2001.

23
Method to Create Categories (pseudocode)

• Assemble exhaustive list of pathogens
• For each pathogen (loop)
• Four experts (adult ID, Ped ID, WMD/ED, Medical
  Informatics) identified the spectrum of outbreak
  size, routes of transmission, indoor outdoor
• Refer to each as a threat
• For each threat (loop)
• Experts asked what kind of automated detection
  system would be needed to detect each threat.
• IF the detection system involved DIFFERENT data
  and algorithm
• create a new category
• ELSE
• add the threat to an existing category
• end

24
List of Detection Problems

• Large scale bioaerosol (e.g., Anthrax)
• Communicable (e.g., SARS)
• Waterborne
• Building contamination
• Foodborne
• Vector borne
• Continuous release
• Single case
• Sexual/blood borne

25
Two Key Patterns of Concern for Detection
Symptoms
Presentation to physician
Deaths
3 4 5 6
3 symptoms pattern 4 nearly
pathognomic 5 culture 6 autopsy
Anthrax or other bioaerosol, food contamination,
water contamination
Contagious disease like Smallpox
Early warning system that collects unorthodox,
nonspecific data and looks for anomalous patterns
Embedded diagnostic expert systems at the point
of care, probably requires POE
Wagner, Dato et al. Data Required for an
Effective Bioterrorism Detection System. Report
to AHRQ, 11/28/01 180 pp.
26
SARS Pattern
27
Detecting SARS- Current Syndromic Approaches
Fever screening Airports, Building entrance,
Hospitals, Schools Testing every sixth
Influenza-like illness ???others
28
Decision Support at Point of Care

• Pretty obvious that it will be needed to do
  better at detecting small outbreaks of unusual
  diseases early
• We already know that it is going to take a long
  time and a lot of money to get there

29
Other SARS Syndromic Strategy(A stretch!(

• Try to monitor for the SARS syndrome
  automatically (SARS is a syndrome, after all)
• Fever
• Respiratory symptoms
• Exposure
• Pneumonia
• Use spatial and other data to try to detect small
  clusters
• In a building
• On a hospital floor
• In a household or family
• In a workplace

30
Individual SARS Case Detection
1. Cough or other respiratory symptom 2.
Temperature gt 38 C 3. Chest x-ray showing
pneumonia or ARDS 4. High risk of exposure
31
1. Respiratory symptom

• NLP of chief complaints to identify patients with
  respiratory syndrome
• Sensitivity 0.77
• Specificity 0.90
• Working on NLP of emergency department reports
• Less timely (1 day after admission)
• But expect better sensitivity/specificity

32
How Well Can We Detect Respiratory Case?
Courtesy Per Gesteland, MD
33
2. Temperature or fever

• Coded temperature (Possibly best, but minority of
  hospitals and temperature may be normal in a
  patient with rigors)
• From NLP of chief complaints
• By NLP of Emergency Department (ED) dictation
• Sensitivity 0.98
• Specificity 0.89
• 1 day delay

34
3. Pneumonia on CXR

• Previous experience with NLP system SymText
  successful
• Keyword search is a reasonable alternative

35
4. Risk of Exposure

• Feasible to obtain on say Air force base
• Less feasible in civilian community but
• Workplace, especially hospital floords
• School
• Buildings
• First three data sources may be sufficient
• Human review of probable SARS cases
• ED reports/ HP exams
• Contact patients

36
Unifying Anthrax and SARS Approach
What we do now for Anthrax
Chief Complaint
Chief Complaint
Chief Complaint
Chief Complaint
Chief Complaint
Chief Complaint
Chief Complaint
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
RespiratorySyndrome
Spatial and temporal analysis to detect
overdensity of cases in a zip code or larger
region
37
Three Slide Overview of The Rest of the Tutorial
38
DATA Which Data Sources are Useful (available
and have an early signal)?
1999 Influenza
Influenza cultures
Sentinel physicians
WebMD queries about cough etc.
School absenteeism
Sales of cough and cold meds
Sales of cough syrup
ER respiratory complaints
ER viral complaints
Influenza-related deaths
Week (1999-2000)
39
ALGORITHMS Which Algorithms Can Extract Maximum
Information from the Data?
Method Has Pitt/CMU tried it? Tried but little used Tried and used Under dev-elopment Multivariate signal tracking? Spatial?
Time-weighted averaging ? ?
Serfling ? ?
ARIMA ? ?
SARIMA External Factors ? ?
Univariate HMM ? ?
Kalman Filter ? ?
Recursive Least Squares ? ?
Support Vector Machine ? ?
Neural Nets ? ?
CuSUM ? ?
Randomization ? ? ?
Spatial Scan Statistics (w/ Howard Burkom) ? ?
Bayesian Networks ? ? ?
Contingency Tables ? ?
Scalar Outlier (SQC)
Multivariate Anomalies ?
Change-point statistics
FDR Tests ? ? ?
WSARE (Recent patterns) ? ? ? ? ?
PANDA (Causal Model) ? ? ? ?
FLUMOD (space/Time HMM) ? ? ?
You will hear about it this from Drs. Moore, and
Hogan
40
Systems The really hard part
National Retail Data Monitor
Detection algorithm
Detection algorithms
RODS
DB
Heavily dependent on existing infrastructure
Hospitals, Retail industry Internet Heavily
dependent on standardization of Architecture,
Security of data, messages
Web
HL7 Listeners
Naive Bayes
DB
VPN
GIS