1. Implementation of Probabilistic Matching in NYC Chronic Hepatitis B and NYC A1C Registries, and Implications Towards an MPI
Maushumi Mavinkurve
Director, Center for Data Matching
NYC Department of Health and Mental Hygiene
October 17th, 2008
Integrated Surveillance Seminar

2. Overview Describe data quality challenges in disease surveillance
Describe probabilistic matching techniques
Implementation of probabilistic matching
NYC Chronic Hepatitis B Registry (LVR)
NYC Hemoglobin A1C Registry (NYCAR)
NYC proposed challenges and benefits of an MPI

3. Public Health Surveillance
Public health surveillance process includes:
Collection of Data on a specific disease or condition via standardized information systems
Analysis and interpretation the data
Dissemination of information to individuals who can act on it
Utilization of information to facilitate necessary response that will effectively deal with the public health issue

4. Surveillance Data Quality Issues
Accuracy
Non-standardized across different data sources
Multiple laboratory systems
De-duplication of reports
Exact duplicates
Multiple events linked to a unique person
Non-relevant information
Accuracy refers to the difference between an estimate of a parameter and its true value.
We characterize the difference in terms of systematic (bias) and random (variance) errors.
Completeness
Integrity
Timeliness refers to the length of time between the reference period of the information
and when we deliver the data product to our customers.
Relevance refers to the degree to which our data products provide information that meets
our customers’ needs.
Accessibility refers to the ease with which customers can identify, obtain, and use the
information in our data products.
Interpretability refers to the availability of documentation to aid customers in
understanding and using our data products. This documentation typically includes: the
underlying concepts; definitions; the methods used to collect, process, and analyze the
data; and the limitations imposed by the methods used.
Transparency refers to providing documentation about the assumptions, methods, and
limitations of a data product to allow qualified third parties to reproduce the information,
unless prevented by confidentiality or other legal constraints.

5. Impact of Data Quality Issues in Surveillance
Impacts on surveillance reporting
Over or underestimates of true cases
Geographical misrepresentation (missing address)
Increases costs
Additional staff required to address data quality issues
Increases inefficiencies
Timeliness for patient or provider follow up

6. Addressing Data Quality Challenges
Modern disease surveillance information systems:
Validates data at time of collection
Minimize inaccurate or incomplete data
Standardizes different data to uniform structure
Integrates matching technology to create
Patient indexes (person-centric systems vs event-centric systems)
Providers indexes
Facility indexes
Could refer to these as each registry as a system that will ultimately feed from a larger MPI.

7. What is Probabilistic Matching?
Rule based match algorithms
Standardizes Data
Parses data into smaller tokens
Create fields that enhance matching
Adapt to specific data - incorporates uniqueness or frequency of data values when comparing records
Processes data in blocks – viable to use on large volume data sets
Rule based match algorithms
employing fields that uniquely identify an entity – name, dob, gender, telephone, etc.
Standardizes Data
Formalizes names: Mike  Michael
Parses data into smaller tokens
Addressline1  house number, street name, street type, apt #
Create fields that enhance matching
Phonetic coding: Soundex, NYSIIS
Hash and packed keys
Adapt to specific data - incorporates uniqueness or frequency of data values when comparing records
“Mary Jones” vs “Maushumi Mavinkurve”
Processes data in blocks – viable to use on large volume data sets

8. Evaluating Match Algorithm
Outcome of a potential match is a weight or likelihood that 2 records are the same entity
Surveillance programs identify thresholds for match algorithm
Prior to reviewing results of match algorithm:
Identify implications for precision (PPV) vs negative predictive valuen (NPV)
Evaluation of health code mandate
Practical issues
Surveillance reporting
Identify guidelines or criteria to review matches

9. Identifying Thresholds
Goal: maximize precision or PPV
Sacrifice on negative predictive value (NPV)
Surveillance programs can decide to review ambiguous matches
Therefore - set high thresholds

10. Outcome of Probabilistic Matching
Entity-centric, relational registry system

11. Background of Hepatitis B in NYC
Decline in acute Hepatitis B incidents case rates (per 100,000 persons) from 11.5 in 1985 to 1.6 in 2006
In NYC burden of chronic Hepatitis B infection as much as 2x higher within specific populations
MSM
IDU
Persons born in regions where HBsAg prevalence >2%
Need for continued surveillance and monitoring
Hepatitis B Surface antigen test was developed and FDA approved in 1980’s
Decline in acute Hepatitis B incidents rates from 11.5 cases per 100,000 persons in 1985 to 1.6 in 2006
In NYC burden of chronic Hepatitis B infection up to 2x higher within specific populations
MSM
IDU
Persons born in regions where HBsAg prevelance >2% (Asian/PI, Eastern Europe, Middle East, Africa, Pacific Island immigrants)
Need for continued surveillance and monitoring
Source: recommendations for identification and public health management of persons with chronic Hepatitis B infection

12. Hepatitis B Surveillance Activities
Monitor disease trends
Aggregate descriptive reporting aimed to guide prevention and intervention efforts
Outreach with newly infected
Educational materials to new cases reported to the registry

13. NYC Hepatitis B Registry
Legacy application, built in-house in 1999
Automatic weekly batch uploads of laboratory reports
Data entry of provider reports
System did not index on patients (event-based), could not link 2 reports for the same person.
Program utilized staff to build and apply deterministic match algorithms
Resource intensive
Version control

14. NYC Liver Virus Registry (LVR)
Implemented in October 2008, built in-house
Migrated all legacy data
Web-based application
Person-centric - integrates probabilistic matching
Consolidated views of all information for a person
Ability to conduct longitudinal analysis
2 weeks ago!
Implemented in October 2008, built in-house
Migrated all legacy data – almost 10 years worth of data
Web-based application
Person-centric - integrates probabilistic matching
Consolidated views of all information for a person
Ability to conduct longitudinal analysis

15. LVR Probabilistic Matching
Created a match algorithm based on fields unique to patient from laboratory and provider reports
Processed all legacy data ~380,000 records
Program evaluated algorithm and identified thresholds
Results: out of ~380,000 reports the match algorithm was able to link these to ~111,000 unique persons
Probabilistic matching enhanced duplication by 1% as compared to legacy deterministic algorithm

16. LVR Challenges & Successes
Iterative review process time and resource intensive
Evaluation against legacy deterministic match
Identifying target PPV and NPV
Successes:
Long term savings on time and resources
Streamlined system
Longitudinal analysis
More accurate case counting
Enhanced data quality
Challenges:
Iterative review process time and resource intensive
Evaluation against legacy deterministic match – NOT GOLD STANDARD, did not evaluate the legacy match
Identifying target PPV and NPV
Successes:
Long term savings on time and resources

17. Implementing Probabilistic Matching with NYC Hemoglobin A1C Registry (NYCAR)

18. What is Diabetes?
Diabetes is a chronic disease caused by inadequate insulin levels or sensitivity leading to elevated blood sugar levels
Blood sugar levels can be measured by
Plasma glucose
Fingerstick glucose
Glycosylated hemoglobin or A1C (goal is <7%)
Persistently high blood sugar levels can cause
Heart disease and stroke
Kidney failure
Blindness
Nerve damage and amputation

19. Diabetes Burden in NYC Diabetes is epidemic in NYC
Prevalence has more than doubled over the past 10 years.
Approximately 500,000 New Yorkers have diabetes
An additional ~200,000 New Yorkers have diabetes, but have not yet been diagnosed
Approximately 1 in 8 adults have diabetes
In 2006, diabetes was the 4th leading cause of death in NYC

20. Prevalence of Self-Reported Diabetes Among Adults in NYC
Source: NYC estimates— CDC Behavioral Risk Factor Survey System (BRFSS) , NYC Community Health Survey (
Source: National estimates—BRFSS 2006

21. Use of Traditional Public Health Surveillance for Chronic Disease
Disease reporting to public health agency to:
Monitor trends
Describe glycemic control in NYC
Identify special populations
Target individuals with poor control
Communicate with provider community
Feedback to providers and their patients
Control epidemics
Decrease complications/improve quality of life

22. Hemoglobin A1C Tests
A1C is a measure of average blood sugar control in preceding 3 months (goal <7%)
A1C is used to:
Monitor individual’s blood sugar control
Guide changes in medication therapy
Impart risk of diabetes complications
Most people who get A1Cs have diabetes so it is a marker for diabetes status
THEREFORE, AN A1C REGISTRY WILL PROVIDE A MECHANISM FOR TRACKING INDIVIDUALS WITH DIABETES
Goal – to have 7.0% (average blood sugar of 170 mg/dL).
THEREFORE, AN A1C REGISTRY WILL PROVIDE A MECHANISM FOR TRACKING INDIVIDUALS WITH DIABETES

23. Implementation of NYCAR
Based on existing NY State / NYC laboratory reporting system
Amendment to NYC health code, Article 13 which mandates communicable disease reporting, to include A1C
Public hearing Summer 2005
Approval of amendment December 2005
Went into effect January 15, 2006
Laboratories submitting data to NY State and NYC subject to mandate
Report information on patient, ordering provider and facility, testing facility and result
Submit via secure network
Receive ~5,000 new lab reports daily – High Volume
Patient advocacy and privacy groups voiced concerns during amendment proceedings
Felt no satisfactory rationale for public health agency involvement in chronic disease reporting
DOHMH clearly wrote into amendment that information can only be released to:
Treating medical provider (s)
Patient
Patients can opt out of intervention but not from being in the registry
Laboratory Reporting:
34 labs reporting A1C tests
Test results reported by lab within 24 hours
PHINMS – Secure file transmission
HL7 messages or ASCII files

24. Objectives of New York City A1C Registry (NYCAR)
Surveillance and epidemiology
Track trends on the population level
Provider feedback and communication
Quarterly provider reports in comparison to peers
Quarterly rosters of patients stratified by A1C level
Patient feedback (via provider)
Letters with A1C information
Local resources
Deliver resources to providers/patients
All of the above requires matching and data linkages
Began January 15, 2006 with mandate of electronic lab reporting.
Provider Reports: Quarterly reports with patients listed by A1C level will be distributed to providers. Reports may be used to identify individuals who may benefit from additional support, such as intensification of therapy, or a referral to a physical activity program or self-management program.
Patient Letters: Letters with recent A1C test results and a reminder to return to care will be sent to patients with high A1C levels.

25. Components of A1C Registry
Information collected by laboratory reports include:
Individual name, address, date of birth, sex
Name and address of ordering provider, ordering facility and testing facility
A1C test collection date and result

26. NYCAR Probabilistic Methodology
Created 3 separate matching models:
Patient
Provider
Ordering Facility
Obtained a representative sample of data
For each model - created a match algorithm utilizing fields that uniquely identify each entity
Name (patient, provider, ordering facility), patient dob, gender, address, providerID, telephone number, etc.
Provided match results to program for review and identify thresholds
Creates indexes for patient, provider and ordering facility:
Each patient appears once and all tests for that individual are linked
Each provider appears once all tests reported by that provider are linked
Each ordering facility appears once and all reports by that facility are linked
Sample used about 100,000 records selected from a specific time period

27. Program Threshold Evaluation
Due to volume of reports, impractical for staff to review all ambiguous matches – need to set thresholds
Method to identify of thresholds using sample
2 reviewers and 1 tie-breaker scored matches referencing guidelines
Utilized a sampling method within weight ranges
Identified specific weight or threshold at which target precision rates were met based on review

28. Deploying Probabilistic Matching
All new incoming A1C lab reports parsed into 3 staging entities:
patient, provider and facilities
Each entity is matched against existing respective entities in the registry
If matched above thresholds, linked to an existing record
If below thresholds, creating a new entity (patient, provider or facility)
Provider Reports and Rosters and Patient Letters are generated using an in-house developed application which reads from the registry
On a weekly basis – the following process occurs

29. Facility Report
Page 2
Note: All information in this slide is fictitious
Page 1

30. Provider Report
Note: All information in this slide is fictitious

31. Patient Letter
GET NEW VERSION SCANNED IN 31

32. Challenges and Successes
Quality of record linkage
Need sufficient information for successful linkage of multiple tests per individual as well as master provider and facility indexing
Maintaining accurate facility-provider linkage
Effect of laboratory variation – availability of data
Review thresholds – time and resource intensive
Successes
Entire process is seamless, electronic and automated
High volume of data
Ability to conduct Longitudinal analysis
Quality of record linkage
Misconceptions – bad quality data cannot be reconciled!
Need sufficient information for successful linkage of multiple tests per individual as well as master provider and facility indexing
Maintaining accurate facility-provider linkage
-1 large umbrella facility can have multiple names
-providers can work for multiple facilities
Case definitions
Individuals with diabetes
Provider for a given patient
Effect of laboratory variation
Impact of inter-laboratory variation – data integrity and availability

33. Is NYC ready for an MPI?

34. NYC Current Status Modernizing several disease registries:
Chronic Hepatitis B - completed
NYCAR – completed
STD – requirements completed
TB – requirements completed
HIV – planning
Is this an opportune time to develop an MPI?

35. Planning an MPI: Challenges
Each registry program has requirements for a matching based on:
Patient population
Data quality and volume
Dissemination/Use of Surveillance data
Foster consensus among disease programs
Breach of Security – higher risk
Legal barriers to creating an MPI
Analysis of health code by reportable disease
Political barriers to creating an MPI
Challenges:
Each registry program has requirements for a matching based on:
Patient population
Data quality and volume
Dissemination/Use of Surveillance data
Foster consensus among disease programs
Legal barriers to creating an MPI
laws, particularly in NYC, are extremely specific. Mandated reportables must be used for the purpose of surveillance and epidemiology of that specific disease – is an MPI a stretch?
Particularly A1C data – this is the most strictly written health code.

36. Planning an MPI: Benefits
Pooling data from different sources could enhance PPV and NPV of the match
Streamline IT resources
Support staff
Infrastructure
Ability to conduct syndemic surveillance and investigation
More efficient use of limited resources
Syndemic is defined as two or more afflictions, interacting synergistically, contributing to excess burden of disease in a population

37. Acknowledgements Diabetes Prevention and Control Program Lynn Silver
Shadi Chamany
Angela Merges
Charlotte Neuhaus
Bahman Tabei
Cindy Driver
Leslie Korenda
Division of Informatics and Information Technology
Don Weiner
Stephen Giannotti
Namrata Kumar
Jisen Ho
Laura Goodman
Bureau of Chronic Disease Control
Katherine Bornschlegl
Magdalena Berger
Emily Lumeng
Division of Epidemiology
Lorna Thorpe
Bonnie Kerker
Jenna Mandel-Ricci
Ram Koppaka

38. Questions? Maushumi Mavinkurve Director, Center for Data Matching
NYC Department of Health and Mental Hygiene
(P)