Transcript No Slide Title

(Pitt)
(CMU)
The Present, Near Term, and
Future of Real-time Public Health
Surveillance
Michael Wagner, MD, PhD
Director, RODS Laboratory
University of Pittsburgh
http://www.health.pitt.edu/rods
The Present
(Using RODS as an example)
Data: ED and Acute Care ADT
Gender
Age
Home Zip
MSH|^~\&||xxx||RODS|200307181731||ADT^A04|2003071XXXX
XXXX|P|2.3<CR>
PID|||||||^05|M|||^^^^15301|||||<CR>
Date and time of registration
PV1||E|||||||||||||||||98765432||||||||||||||||||||||
||200307181731||<CR>
DG1||||CARBON MONOXIDE EXPOSURE<CR>
IN1||||||||||||||||||||||||||||||||||||||||||||^^^^<C
Chief complaint
R>
HL7 admission/discharge/transfer message
about a patient registration in an emergency
department
Data: OTC Sales
• ~7500 products (UPC codes) used for selftreatment of infectious diseases
• 18 analytic classes at present (“categories”)
Cold Relief Adult Liquid (709 products)
Cold Relief Adult Tablet (2467)
Cold Relief Pediatric Liquid (323)
Cold Relief Pediatric Tablet (74)
Cough Syrup Adult Liquid (592)
Cough Syrup Adult Tablet (32)
Cough Syrup Pediatric Liquid (24)
Nasal Product Internal (371)
Throat Lozenges (364)
Antifever Pediatric (274)
Antifever Adult (1340)
Bronchial Remedies (43)
Chest Rubs (78)
Diarrhea Remedies (165)
Electrolytes Pediatric (75)
Hydrocortisones (185)
Thermometer Pediatric (125)
Thermometer Adult (313)
Numbers in parenthesis are the number of UPC codes in the category
Analysis: What we do now to detect
Influenza, Crypto, Anthrax
Chief
Chief
Chief
Chief
Complaint
Complaint
Complaint
Complaint
Chief
Chief
Chief
UPC
Chief
code
Complaint
Complaint
Complaint
Complaint
Bayes
Classifier
Respiratory
Respiratory
Respiratory
Respiratory
Syndrome
Syndrome
Syndrome
Syndrome
Category
mapping
Respiratory
Respiratory
Respiratory
Respiratory
Pediatric
Syndrome
Syndrome
Syndrome
Electrolyte
Syndrome
Univariate
BARD
Spatial WSARE
Scan
Algorithms that perform spatial
and temporal analysis to detect
overdensities of cases in a zip
codes or larger regions
Spatial Scanning of Electrolyte Sales
• Ultra fast version
typically turns multiday analysis into less
than an hour.
• Searches all
rectangular regions.
Role in surveillance:
find clusters of
illness with shapes
such as oriented
rectangles.
More info: http://www.autonlab.org/autonweb/showProject/4/
BARD (Bayesian Aerosol Release Detector)
Approach
• BARD automates the analysis done
by Messelson et al of the
meteorological conditions in
Sverdlovsk in the days prior to the
outbreak.
• The method uses a computational
inversion of the well-known
Gaussian dispersion model,
although any dispersion model can
be used.
Role in surveillance: detect
patterns of disease activity
consistent with the weather
and the incubation time of
Anthrax (or other organism
that can be released as an
aerosol)
: Meselson et al, 1994 Science
More info: BARD Tech report
WSARE (“What’s Strange About Recent Events”)
Approach
REPRESENTATIVE SURVEILLANCE DATA
Date Time Hospit ICD9 Prodrom Gender Age Home
Many
al
e
Location more…
6/1/03 9:12
1
781
Fever
M
20s
NE
…
6/1/03 9:45
1
787
Diarrhea
F
40s
SE
…
:
:
:
:
:
:
:
:
:
WSARE Approach
Standard Approach
Select in advance which
subpopulations to monitor
(e.g., each county, zip)
Monitor hundreds of
thousands of
subpopulations
Do not pay close attention
to effect of multiple testing
Pay close attention to
effect of multiple testing
• Search over hundreds of thousands of subpopulations
• For each subpopulation, use as good of a model as can
be created to predict expected counts
• Compute p value, taking into account multiple testing
All
Historical
Data
Today’s
Environment
What
should be
happening
today?
Today’s
Cases
What’s strange about
today, considering its
environment? And how
significant is this?
Evaluation
Detailed comparison on 2,000 simulated scenarios and
Western PA ED Data
Role in Surveillance
Standard
WSARE2.0
Detect small clusters of illness in
healthcare workers, age groups,
workplaces …
WSARE2.5
WSARE3.0
More info: http://www.autonlab.org/autonweb/showProject/4/
Evaluations
Performance of Bayesian parser
Sens.
Spec.
Respiratory
5
60-77%
90-94%
GI
4
63-74%
90-96%
Neurological
3
68-72%
93-95%
Rash
3
47-60%
99%
Botulinic
3
17-30%
99%
Hemorrhagic
1
75%
98%
Constitutional
1
46%
97%
Example of a Detectability Analysis
Public Water Drinking Advisory
Antidiarrheal Sales
Number
Studies
Syndrome
Case Study of OTC Monitoring
Sensitivity
Case studies
10%
1%
0.1%
Affected
Days into outbreak
RODS Deployments
Jurisdiction
Hospitals (real
time/total)
Platforms
Pennsylvania
50/53
Utah
27/27
Ohio
10/13
Solaris/
Oracle
(All these jurisdictions are
using the Pittsburgh server
facilities)
Atlantic City, NJ
Michigan
Taiwan
Houston, TX
El Paso, TX
Los Angeles, CA
Mississippi (OTC)
3/3
2500 visits/day
Unix/Oracle
0/190+
Unix/Oracle
1/1, 2 more pending
Aug ‘04
Windows/Oracle
6 pending Aug ‘04
Windows/MSSQL
pending
Windows/MSSQL
(1 from U Miss)
Linux/Oracle
The Present (summarized)
• Data
• More widely collected
• Real-time HL7 transmission of chief complaints from emergency departments
• OTC data with 12-24 hour delay
• Web based disease reporting
• Less widely collected
• Call center
• Electronic lab reporting
• School absenteeism
• Algorithms
•
•
•
•
Univariate
Multivariate
Spatial scanning
Aerosol detection
• Deployments (not just RODS)
• Many jurisdictions
• Spotty in terms of data coverage
• Evaluation
• Understanding of methodology is good
• Understanding of detectability in its infancy
The Near Term (next two
years)
More Data Means Enables More Specific Case
Detection
What we do now
Chief
Chief
Chief
Chief
Complaint
Complaint
Complaint
Complaint
Respiratory
Respiratory
Respiratory
Respiratory
Syndrome
Syndrome
Syndrome
Syndrome
Spatial and temporal
analysis to detect
overdensity of cases in a
zip code or larger region
Future
Chief
Chief
Chief
Pneumonia
Chief
Complaint
Chief
Complaint
Chief
Complaint
Chief
Complaint
on X-ray Complaint
Chief
Complaint
Complaint
Complaint
Chief
Chief
Chief
Temperature
Chief
Complaint
Complaint
Complaint
Complaint
Respiratory
Respiratory
Respiratory
Respiratory
SARS
Syndrome
Syndrome
Syndrome
Syndrome
Spatial and temporal
analysis to detect small
number of cases in a
hospital or hotel
One Year of Daily Counts of Febrile Illness in
Beijing
The Hospital Message Router*
Electronic
Health
Record
Lab
Radiology
Message
Router
Scheduling/
Registration
Transcription
Billing
Pharmacy
Hospital IT Infrastructure
*aka
Interface Engine aka Integration Engine
Algorithms that Process Multiple Data
Streams
Red: Cough Sales
Signal
Blue: ED Respiratory Visits
Cough Sales
This is an
anomaly
One Sigma
2 Sigma
ED Respiratory Visits
Near term summarized (next 2 years)
• Data
• Clinical: Laboratory results (Micro), radiology,
temperatures, outpatient data military
• Integration of data from Biowatch sensors and military
healthcare
• Algorithms
• That process multiple data streams
• That ntegrate water supply routes into surveillance
• For building monitoring
• More complete deployments in major cities and
across border
• Evaluations of detectability, studies of behavior of
sick individuals
The Future
Approaching the theoretical limits of detectability
(first case on day of infection ): Detecting small
outbreaks and detecting bigger outbreaks earlier
• Through better surveillance data
• More an better biosensors
• Earlier detection of patients with fever, constitutional symptoms
• Passive—embedded chips, smart toilets,
• Active—self reporting
• Widespread use of microchip arrays for diagnosis
• Environmental and intelligence data to set priors
• Decision support at the point of care
• … and algorithms that can extract all information from
knowledge and data
• Social networks
• Food distribution information
• …
PANDA 2
Approach
Aggregate
Observations
 PANDA2 is designed to be able to fuse
ALL data and knowledge to achieve the
very earliest detection.
One Patient
 The method involves using causal
networks that models the relationships
between causes of disease outbreaks and
the health state of individuals in the
population, as well as aggregate
observations about the population (e.g.,
levels of OTC sales)
 These causal networks represent prior
knowledge about disease presentation as
well as knowledge about the spatio
temporal spread of outbreaks
A city-wide model
More info: Uncertainty in AI paper