1. Patient Matching Algorithm Challenge
Informational Webinar
Caitlin Ryan, PMP| IRIS Health Solutions LLC, Contract Support to ONC
Adam Culbertson, M.S., M.S. | HIMSS Innovator in Residence, ONC

2. The Current State of Patient Matching Challenge
Agenda
ONC Overview
The Current State of Patient Matching
Challenge
Overview
Registration
Project Submissions
Winners and Prizes
Eligibility Requirements
Q&A

3. Office of the National Coordinator for Health IT (ONC)
The Office of the National Coordinator for Health Information Technology (ONC) is at the forefront of the administration’s health IT efforts and is a resource to the entire health system to support the adoption of health information technology and the promotion of nationwide health information exchange to improve health care,
ONC is organizationally located within the Office of the Secretary for the U.S. Department of Health and Human Services (HHS),
ONC is the principal federal entity charged with coordination of nationwide efforts to implement and use the most advanced health information technology and the electronic exchange of health information.

4. ONC Challenges Overview
The statutory authority for Challenges are hosted under Section of the America COMPETES Reauthorization Act of (Public L. No ).
ONC Tech Lab - Innovation
Spotlight areas of high interest to ONC and HHS
Direct attention to new market opportunities
Continue work with start-up community and administer challenge contests
Increase awareness and uptake of new standards and data

5. Patient Matching Definition
Patient matching: Comparing data from multiple sources to identify records that represent the same patient. Also called merge-purge, record linkage and entity resolution in other fields.
Entity Resolution, Data Matching, Record Linkage, Object Identification, Merge-Purge, Linked Data, Data Deduplication, Identify Resolution, Schema Matching

6. Source: Culbertson, A. Patient Matching A-Z, Wednesday, March 2ndst HIMSS 2016, Las Vegas, NV

7. Significant Dates in (Patient) Matching
Campbell, K et al
A Comparison of Link Plus,
The Link King, and a “Basic”
Deterministic Algorithm
A Framework for
Cross-Organizational
Patient Identity
Management
HIMSS
Patient Identify
Integrity Toolkit,
Patient Key
Performance
Indicators
RAND Health Report
Identity Crisis: An
Examination of the
Costs and Benefits
of a Unique Patient
Identifier for the US
Health Care System
2008
2015
Kho, Abel N., et al
Design and
Implementation of
a Privacy Preserving
Electronic Health
Record Linkage Tool
Dunn
Record
Linkage
1946
Fellegi & Sunter
A Theory of
Record Linkage
1969
Winkler
Matching and
Record Linkage
2011
Soundex
US Patent
1918
Newcombe,
Kennedy, & Axford
Automatic Linkage
of Vital Records
1959
Grannis, et al
Analysis of Identifier
Performance Using a
Deterministic Linkage
Algorithm
2002
Grannis, et al
Privacy and Security
Solutions for Interoperable
Health Information
Exchange
2009
Audacious Inquiry
and ONC
Patient Identification and
Matching Final Report
2014
Joffe et al
A Benchmark Comparison
of Deterministic and
Probabilistic Methods for
Defining Manual Review
Datasets in Duplicate
Records Reconciliation
HIMSS
Patient Identity
Integrity
Exemplars
Intelligence Industry post 9/11 realized that had a problem matching non-Anglo names
Problem faced in many verticals such intelligence finance, marketing
A Comparison of Link Plus, The Link King, and a “Basic” Deterministic Algorithm Campbell, K et al. (2008) (Notes: Campbell, Kevin M., Dennis Deck, and Antoinette Krupski. "Record linkage software in the public domain: a comparison of Link Plus, The Link King, and a basic' deterministic algorithm." Health Informatics Journal 14.1 (2008): 5-15.)
A benchmark comparison of deterministic and probabilistic methods for defining manual review datasets in duplicate records reconciliation Joffe et al. Jan 14.
 (Notes: Joffe E, Byrne MJ, Reeder P, et al. A benchmark comparison of deterministic and probabilistic methods for defining manual review datasets in duplicate records reconciliation. Journal of the American Medical Informatics Association : JAMIA. 2014;21(1): doi: /amiajnl )
“work of the IIR will lead to establishing metrics of patient matching technology approaches and create a pathway for evaluating solutions.”   and " IIR to develop a vision, strategy, and implementation plan for the near-term deployment of consistent patient data matching in health that builds on the body of work from HHS’s Office of the National Coordinator for Health IT (ONC) and healthcare community partners.  The IIR will also assess the longer-term applicability of identity management methods, processes and technologies currently in use in healthcare and other sectors.” 
Design and implementation of a privacy preserving electronic health record linkage tool in Chicago." Journal of the American Medical Informatics Association (2015): ocv038.
Dusetzina, Stacie B., et al
Linking Data for Health
Services Research: A Framework
and Instructional Guide
HIMSS hires Innovator
In Residence (IIR) focused
on Patient Matching
Source: Culbertson, A. & Miller, K., Patient Matching EHR Ailments: Going from Placebo to Cure, Tuesday, March 1st HIMSS 2016 Las Vegas, NV

8. Data quality issues make matching more complicated
Problem
Patient data matching has been noted as one key barriers to achieving interoperability in the Nations Road Map for Health IT
Patient Matching causes issues for over 50% of health information managers1
Problem will increases as we increase the volume of health data sharing
Data quality issues make matching more complicated
Lack of knowledge of patient matching algorithms performance, adoption of metrics
1) MMS - all known milestones
ONC Annual Plan – subset of database focused on ONC ownership incremented into quarterly milestones
ONC Quarterly Dashboard – subset of annual plan for quarterly monitoring and cascade into NC & SES perforamnce plans
ONC Bi-Weekly / Monthly Milestone monitoring report 1)

9. Garbage in and Garbage out
Data Quality
Data Quality is a Key
Garbage in and Garbage out
Data entry errors are compound data matching complexity
Various algorithmic solutions to address these, not perfect
Types of errors:
Missing or Incomplete Values
Inaccurate data
Fat finger errors
Information is out of date
Transposed names
Misspelled names

10. The 5 Step Data Match Process
Data pre-processing
Indexing
Comparison
Classification
Evaluation
Pre-processing: characterizes the data and ensures the elements have the same structure and the content follows the same format
Indexing: organizing the data to support better pairing (blocking and the use of a blocking key common)
Comparison: identifying the similarity between two records seeking comparison vectors
Classification: based on comparison results, records are found to be matches, non-matches or potential matches
Data pre-processing: 3 to 4 steps:
Remove unwanted characters and words
Expand abbreviations and correct misspellings
Segment attributes into well-defined and consistent output attributes
Verify the correctness of attribute values
Important to note that this step must not overwrite the original input data – but new attributes can and should be created that contain the cleaned and standardized data.
Evaluation: comparing match results with the known ground truth or gold standard.
Data Matching; Concepts and techniques for record linkage, entity resolutions, and duplication detection, Peter Christen 2012

11. "If you can't measure it, you can't improve it.“
Solution
"If you can't measure it, you can't improve it.“
-Peter Drucker

12. ONC is launching the Patient Matching Algorithm Challenge!
Solution
ONC is launching the Patient Matching Algorithm Challenge!

13. Metrics for Algorithm Performance

14. Ideal outcome of any matching exercise is correctly answering this one question hundreds or thousands of times, Are these two things the same thing?
Correctly identifying all the true positives and true negatives while minimizing the number of errors, false positives and false negatives
Patient Matching Goal
Source: Culbertson, A. Patient Matching A-Z, Wednesday, March 2ndst HIMSS 2016, Las Vegas, NV

15. Patient Matching Terminology
True Positive- The two records represent the same patient
True Negative- The two records don't represent the same patient
Source: Culbertson, A. Patient Matching A-Z, Wednesday, March 2ndst HIMSS 2016, Las Vegas, NV

16. Patient Matching Terminology
False Negative: The algorithm misses a record that should be matched
False Positive: The algorithm creates a link to two records that don’t actually match
Source: Culbertson, A. & Miller, K., Patient Matching EHR Ailments: Going from Placebo to Cure, Tuesday, March 1st HIMSS 2016 Las Vegas, NV

17. Evaluation EHR A EHR B Truth (Gold Standard) Algorithm Match Type
Jonathan
True Positive
Sally
Non-Match
True Negative
False Positive
Jon
False Negative
Good
Bad
Add color to highligh differences
Source: Culbertson, A. Patient Matching A-Z, Wednesday, March 2ndst HIMSS 2016, Las Vegas, NV

18. Evaluation EHR A EHR B Truth (Gold Standard) Algorithm Match Type
Jonathan
True Positive
Sally
Non-Match
True Negative
False Positive
Jon
False Negative
Bad
Add color to highligh differences
Source: Culbertson, A. Patient Matching A-Z, Wednesday, March 2ndst HIMSS 2016, Las Vegas, NV

19. Evaluation EHR A EHR B Truth (Gold Standard) Algorithm Match Type
Jonathan
True Positive
Sally
Non-Match
True Negative
False Positive
Jon
False Negative
Bad
Add color to highlight differences
Source: Culbertson, A. Patient Matching A-Z, Wednesday, March 2ndst HIMSS 2016, Las Vegas, NV

20. Precision = True Positives / (True Positives + False Positives)
Evaluation
Truth
Algorithm
Positive
Negative
True Positive
False Positive
False Negative
True Negative
Precision
Recall = True Positives / (True Positives + False Negatives)
Recall
Source: Culbertson, A. Patient Matching A-Z, Wednesday, March 2ndst HIMSS 2016, Las Vegas, NV

21. Tradeoffs between Precision and Recall
Evaluation
Calculation
Precision
Recall
Tradeoffs between Precision and Recall
F-Score is the harmonic mean between precision and recall
F score. =
Precision
If Jon Smith is the record you are looking for and you search for Jon Smith. Precision is the likelihood that the record you have that corresponds to john smith is actually Jon Smith
Recall – If Bob Jones has 100 records in a database how many of Bob Jones records to you get? If you get 90 Recall you got 90 of bob Jones Records
Different use cases have different requirements, and different balance of tradeoffs. Security vs Healthcare

22. Creating Test Data Sets

23. Development of Test Data Set
Patient Database
Select Potential Matches
(aka Adjudication Pool)
Manual
Reviewer 1
Reviewer 2
Reviewer 3
Human-Reviewed Match Decisions
(Answer Key == Ground Truth Data Set)
MITRE help in graphics
Compare Algorithm and
Test Data Set
Source: Culbertson, A. & Miller, K., Patient Matching EHR Ailments: Going from Placebo to Cure, Tuesday, March 1st HIMSS 2016 Las Vegas, NV

24. Development of Ground Truth Sets
Identify data set that reflects real word use case
Develop potential duplicates
Human adjudication review and classification
Match or Non-Match
Estimate truth
Pooled methods using multiple matching methods

25. Issues In Establishing Ground Truth
First step in evaluation is to determine why the evaluation is being conducted
Different truth for different applications
Security Applications vs Patient Health Record
What is the cost of missing a match?
Security: Lives are lost
Health: Patient safety event, missed medications, allergies, etc… death But…this is situation today.
What is the cost of wrongly identifying a match?
Security : Passenger is inconvenienced / delayed
Health: Patient safety event, wrong medication, treatment, liability, death
Criteria for truth must be carefully established and well- understood
E.g. Question posed to annotators must be carefully phrased
Summary for Healthcare Use Case

26. Issues In Establishing Ground Truth (cont’d)
Different truth for different applications
Credit check
Security applications
Customer support
De-duplication of mailing lists
What is the cost of missing a match?
New record entered into database
Irritated customer
Lives are lost
Criteria for truth must be carefully established and well-understood by annotators
Question posed to annotators must be carefully phrased

27. Issues In Establishing Ground Truth (cont’d)
How much time / expertise is available to judge (/discount) false positives?
Needs to reflect real word test use case
Evaluation results are only as good as the truth on which they are based
And only as appropriate as the evaluation is to the task that will be performed with the operational system
Absolute recall impossible to measure without completely known test set (i.e. “You don’t know what you’re missing.”)
Estimate with pooled results

28. B Smith  Bill Smythe  William Smythe  W Smith ??
Examples
B Smith  Bill Smythe  William Smythe  W Smith ??
DOB: 10/12/1972  October 11, 1972 
December 10, 1972  12/10/72  October 12, 1927
Use all same name example throughout

29. Creating Ground Truth Answer Key
1 John Smith Washington, DC 2 Carol Jones Bethesda, MD 3 Bobby Johnson Arlington, VA 4 John Smith Washington, DC
1 John Smith Washington, DC
4 Johnny Smith Washington, DC Carol Jones Bethesda, MD 3 Bobby Johnson Arlington, VA

30. Known potential duplicate pairs mimic real world scenarios
Challenge Dataset
Dataset synthetically generated by Just Associates using a proprietary software algorithm
 Based on real-world data in an MPI and actual data discrepancies in each field across the fields
 Known potential duplicate pairs mimic real world scenarios
  Does not contain PHI
  Approximately 1M Patient Records 
Fields include:
Enterprise ID, LAST NAME, FIRST NAME, MIDDLE NAME, SUFFIX, DOB, GENDER, SSN, ADDRESS1, ADDRESS2, CITY, STATE, ZIP, PHONE, PHONE2, ,   ALIAS, MOTHERS_MAIDEN_NAME, MRN, SSNs(Most are within the 800 range)
CSV format

31. Patient Matching Algorithm Challenge
Contestants will unlock a test data set provided by ONC
This data set must be used to run algorithms and provide their results for evaluation.
Submissions will be scored against an “answer key”
Receive performance scores and appear on a Challenge leaderboard.
Up to six participants will be selected as winners for this Challenge and awarded cash prizes
Website:
Participants will unlock and download the test data set at the time of registration. Participants will then run their algorithms and submit their results to the scoring server on the Challenge Website.
A small set of true match pairs (that have been created and verified through manual review) exist within the large data set and will serve as the “answer key” against which participants’ submissions will be scored.

32. Eligibility Requirements
Shall have registered to participate in the Challenge under the challenge requirements by ONC.
Shall have complied with all the stated requirements of the Challenge.
Businesses must be incorporated in and maintained a primary place of business in the United States; Individuals must be a citizen or permanent resident of the United States.
Shall not be an HHS employee.
All individual members of a team must meet the eligibility requirements.

33. Eligibility Requirements (cont’d)
May not be a federal entity or federal employee acting within the scope of their employment.
Federal grantees may not use federal funds to develop COMPETES Act challenge applications unless consistent with the purpose of their grant award.
Federal contractors may not use federal funds from a contract to develop COMPETES Act challenge applications or to fund efforts in support of a COMPETES Act challenge submission.
Participants must also agree to indemnify the Federal Government against third party claims for damages arising from or related to Challenge activities.

34. Registration is NOW open!
Visit the challenge website and fill in all required fields of the registration form
Create a username and password and team name which will be used on the leader board

35. The submission period* will be open for 3 months
Project Submissions
The submission period* will be open for 3 months
100 submissions from each individual/team
Can submit at any time during the “submission period”
Submissions will be allowed until 11:59pm on the last day of the submission period
*Submission period dates have not been determined. Once the test data set is available
we will add these dates to the challenge website.

36. Leader Board Example
Keith J. Miller, The MITRE Corporation , Biographic (Matching) Technologies, GIS, Tampa, FL September 21, 2015

37. Leader Board Example

38. Precision Recall Best first f-score run
Winners and Prizes
The Total Prize Purse for this challenge is $75,000
Judging will be based upon the empirical evaluation of the performance of the algorithms.
Highest F-Score
1st- $25,000
2nd- $20,000
3rd- $15,000
Best in Category: ($5,000 ea. Category)
Precision
Recall
Best first f-score run
Judging will be based upon the empirical evaluation of the performance of the algorithms. For the purposes of scoring, all scores will be assessed to three places after the decimal (i.e., 0.xxx).
F-score judging and award conditions:
In general, the top three algorithms with the highest F-scores will be selected as the winners.
In the event that a team achieves F-scores that would consecutively occupy 1st, 2nd, and/or 3rd place (or any consecutive combination or otherwise (e.g., (1st and 3rd) or (2nd & 3rd)), the team will be awarded only one prize, which will be the highest available prize. After that award, that teams remaining score(s) will be skipped until a different competitor can be awarded the next prize level.
In the event of a tie in F-scores, the winners will be awarded their prizes based on least amount of “runs”/tries that it took to get such score. For example, if two teams tied for first place and one got its high F-score on its 31st try and the other team on its 57th try, the first team would be awarded 1st place and the other team 2nd place. The same tie-breaking approach will be applied for all award positions. If the least amount of tries method subsequently results in a tie, the teams will be awarded that place and the prize will be split evenly to all who won that place.
Best in category prizes
In general, best in category prizes will be made for precision, recall, and best first run.
Unlike the F-score award rules, a team is permitted to win one or all of these “best of” category prizes.
In the event that a tie occurs for the precision and recall categories, the same least amount of tries method (as noted in F-score) will be used to break the tie and same even award split method will be used if the least amount of tries method does not break the tie.
If the best first F-score run category results in a tie, the award will be split evenly to all who won the category.

39. Webinars on how to participate and challenge overview
Get Involved
Webinars on how to participate and challenge overview
May 17th
May 24th
Kicking off Patient Data Matching Algorithm Challenge in June
Website:

40. Acknowledgments
Thank you to the following individuals and organizations for their involvement in the planning and development of this challenge :
Debbie Bucci and the ONC Team
HIMSS North America, Tom Leary, HIMSS
Greg Downing, HHS Idea Lab, ONC
Jerry and Beth Just and the Just Associates Team
Keith Miller and Andy Gregorowicz, MITRE
Caitlin Ryan, IRIS Health Solutions
Capitol Consulting Corporation Team

41. FOR ADDITIONAL QUESTIONS/INFORMATION CONTACT:
Adam Culbertson,
Debbie Bucci, (preferred)
Phone:

42. Thank you for your interest!
The ONC Team