Page 1
Syndromic Surveillance in Montreal: An Overview of Practice and Research
David Buckeridge, MD PhD
Epidemiology and Biostatistics, McGill University
Surveillance Team, Montreal Public Health
QPHI Surveillance Meeting
KFL&A Public Health, Kingston, ON
June 13th, 2008
Page 2
Syndromic Surveillance in Montreal (ou, Vigie Multirisque)
Population Under Surveillance
Intervention Decision
Intervention
GuidelinesPublic Health Action
3. Conveying information for action
Population Pattern
Definitions
Pattern Report
Pattern Detection Algorithm
2. Detecting population patterns
Event Report
s
Individual Event Definitions
Event Detection Algorithm
Data Describing Population
1. Identifying individual cases
Decision Algorithm
Knowledge
Telehealth911 CallsHospitalReportable
Telehealth911 CallsHospitalReportable
Counts, Native coding schemes, ISDS consensus syndromes
Counts, Native coding schemes, ISDS consensus syndromes
Routine SaTScan, alerts for shared addresses
Routine SaTScan, alerts for shared addresses
Daily review of analysis results, not clear protocol
Daily review of analysis results, not clear protocol
Page 3
Vigie Multirisque: Data Sources
Emergency Departments Currently: All 22 ED in Montreal via web form, total
counts, no diagnosis or chief complaint Future: Automated feeds under development, triage
code and level, chief complaint, postal code
EMS Dispatch and Billing Long-Term Care Tele Health Reportable Diseases
Page 4
Vigie Multirisque: Dashboard
Page 5
Vigie Multirisque: Analysis
Page 6
Vigie Multirisque: Analysis
Page 7
Vigie Multirisque: Descriptive
Page 8
Surveillance Research
Page 9
Syndromic Surveillance Research
Population Under Surveillance
Intervention Decision
Intervention
GuidelinesPublic Health Action
3. Conveying information for action
Population Pattern
Definitions
Pattern Report
Pattern Detection Algorithm
2. Detecting population patterns
Event Report
s
Individual Event Definitions
Event Detection Algorithm
Data Describing Population
1. Identifying individual cases
Decision Algorithm
Knowledge
Subsets of admin data for ILI surveillance
Subsets of admin data for ILI surveillance
Page 10
Looking for the Leading ILI Indicator in Billing Data
Page 11
Syndromic Surveillance Research
Population Under Surveillance
Intervention Decision
Intervention
GuidelinesPublic Health Action
3. Conveying information for action
Population Pattern
Definitions
Pattern Report
Pattern Detection Algorithm
2. Detecting population patterns
Event Report
s
Individual Event Definitions
Event Detection Algorithm
Data Describing Population
1. Identifying individual cases
Decision Algorithm
Knowledge
Subsets of admin data for ILI surveillance
Subsets of admin data for ILI surveillance
Accuracy of ICD codes and syndromes in ambulatory practice
Accuracy of ICD codes and syndromes in ambulatory practice
Page 12
Syndromic Surveillance Research
Population Under Surveillance
Intervention Decision
Intervention
GuidelinesPublic Health Action
3. Conveying information for action
Population Pattern
Definitions
Pattern Report
Pattern Detection Algorithm
2. Detecting population patterns
Event Report
s
Individual Event Definitions
Event Detection Algorithm
Data Describing Population
1. Identifying individual cases
Decision Algorithm
Knowledge
Subsets of admin data for ILI surveillance
Subsets of admin data for ILI surveillance
Accuracy of ICD codes and syndromes in ambulatory practice
Accuracy of ICD codes and syndromes in ambulatory practice
1. Selecting the best algorithm2.
3.
1. Selecting the best algorithm2.
3.
Page 13
Building the Knowledge-Base for Algorithm Selection
1. Model the aberrancy detection process
2. Evaluate modeled algorithms using high throughput software
3. Use machine learning to identify and model the determinants of detection
Page 14
Syndromic Surveillance Research
Population Under Surveillance
Intervention Decision
Intervention
GuidelinesPublic Health Action
3. Conveying information for action
Population Pattern
Definitions
Pattern Report
Pattern Detection Algorithm
2. Detecting population patterns
Event Report
s
Individual Event Definitions
Event Detection Algorithm
Data Describing Population
1. Identifying individual cases
Decision Algorithm
Knowledge
Subsets of admin data for ILI surveillance
Subsets of admin data for ILI surveillance
Accuracy of ICD codes and syndromes in ambulatory practice
Accuracy of ICD codes and syndromes in ambulatory practice
1. Selecting the best algorithm2. Looking for connected cases3.
1. Selecting the best algorithm2. Looking for connected cases3.
Page 15
System Architecture
Python, R-Server, SaTScan
Statistical Analysis Server
PostGreSQL / PostGIS DB
Spatial Database
Apache + PHP, MapServer +
MapScript
Mapping and Web
ServerWeb Client
Firefox, Explorer
Current Case Management System
DCIMI Database
DCIMI Client
Oracle
Oracle Forms
Web-based Cartography Software
Page 16
Organizing Data by Person, Place and Time
PostGreSQL / PostGIS DB
Spatial Database
PersonMADONameBirthdate…
PlaceAddressX, YPlace Type (Residence, Workplace)…
EpisodeOnset DateDisease Type…
Contact
SituationRole (Home, Work, School, …)Active Date…
Page 17
Address Validation and Correction in a Public Health System
Page 18
Dracones – Query Form
PersonPerson
TimeTime
PlacePlace
Page 20
Dracones – SaTScan Results
Page 21
Syndromic Surveillance Research
Population Under Surveillance
Intervention Decision
Intervention
GuidelinesPublic Health Action
3. Conveying information for action
Population Pattern
Definitions
Pattern Report
Pattern Detection Algorithm
2. Detecting population patterns
Event Report
s
Individual Event Definitions
Event Detection Algorithm
Data Describing Population
1. Identifying individual cases
Decision Algorithm
Knowledge
Optimal decision making after an alarm
Optimal decision making after an alarm
Subsets of admin data for ILI surveillance
Subsets of admin data for ILI surveillance
Accuracy of ICD codes and syndromes in ambulatory practice
Accuracy of ICD codes and syndromes in ambulatory practice
1. Selecting the best algorithm2. Looking for connected cases3. Spatial TB clusters
1. Selecting the best algorithm2. Looking for connected cases3. Spatial TB clusters
Page 22
Using Surveillance Information to Manage Outbreaks Effectively
Much research on the statistical accuracy of aberrancy detection algorithms
Little attention to what happens next Some attempts to describe response protocols (e.g.,
flow chart, wait a day) No quantitative modeling of response
Rational response is important Small window to obtain benefit Surveillance information uncertain
Page 23
The Traditional Surveillance Alert Response Model
Environmental DataKnowledgeKnowledge
Detection Method
Detection Method
Investigate
Investigate
ConfirmConfirm
WaitWait
Review RecordsReview Records No
OutbreakNo
Outbreak
Alert No AlertInterventionNo
Intervention
Yes
Yes
Yes
No
No
No
Page 24
Identifying an Optimal Policy
The goal is to identify a policy, or a mapping from a belief state (probability distribution over states) to actions
The belief state, provides the same information as maintaining the complete history
Value iteration is used to solve POMDP
Page 25
Applying a POMDP to Surveillance
S - True outbreak state {No Outbreak, D1, ….}
O - Output from detection algorithm {0,1}
A - Possible public health actions
T(s,a,s’) - Impact of actions given the state
R(s,a) - Costs of actions and outbreak states
Do nothing
Review
records
Investigate
cases
Declare
outbreak
Actio
n
Transitio
n
(Izadi M & Buckeridge DL, 2007)
Page 26
POMDP Policy Dominates Ad Hoc Policy
Page 27
Syndromic Surveillance Research
Population Under Surveillance
Intervention Decision
Intervention
GuidelinesPublic Health Action
3. Conveying information for action
Population Pattern
Definitions
Pattern Report
Pattern Detection Algorithm
2. Detecting population patterns
Event Report
s
Individual Event Definitions
Event Detection Algorithm
Data Describing Population
1. Identifying individual cases
Decision Algorithm
Knowledge
Optimal decision making after an alarm
Optimal decision making after an alarm
Evaluating Syndromic Surveillance in Public Health Practice: Detecting Waterborne Outbreaks
Evaluating Syndromic Surveillance in Public Health Practice: Detecting Waterborne Outbreaks
Subsets of admin data for ILI surveillance
Subsets of admin data for ILI surveillance
Accuracy of ICD codes and syndromes in ambulatory practice
Accuracy of ICD codes and syndromes in ambulatory practice
1. Selecting the best algorithm2. Looking for connected cases3. Spatial TB clusters
1. Selecting the best algorithm2. Looking for connected cases3. Spatial TB clusters
Page 28
Automated and ‘Traditional’ Surveillance for Waterborne Outbreaks
Dispersion Exposure
Latent InfectedLatent
Infected
Infectious (Symptom
atic)
Infectious (Symptom
atic)
Infectious (Asympto
matic)
Infectious (Asympto
matic)
Tele-healthTele-health
EDED
Out-patientOut-
patient
Stool TestStool Test
Analysis by Public
Health
Analysis by Public
Health
Disease Health Care Utilization Reportable Disease Surveillance
Syndromic Surveillance
Analysis by Public
Health
Analysis by Public
Health
Outbreak DetectionOutbreak Detection
Historical Tele-health and ED Data
Historical Case Reports
O
S
S
S
R
R
RR
R
O O S,R
Page 29
Modeling Dispersion of Microorganisms
Dispersion
Page 30
Modeling Infection: Mobility
Page 31
Mobility-Weighted Infection Probability by Home Address
Page 32
Modeling Disease, Visits, Testing, Reporting to Public Health
Page 33
Evaluating the Effect of Surveillance Enhancements
Page 34
For more information…