Patient identity resolution in healthcare analytics is the process of determining whether records from different systems refer to the same person and then linking those records into a coherent identity. The HL7 Interoperable Digital Identity and Patient Matching implementation guide extends the FHIR Patient $match operation for cross-organizational use through a profiled operation called $IDI-match, and conformant systems must support Patient $match or an equivalent non-FHIR capability [1][2]. The Office of the National Coordinator for Health Information Technology defines patient matching more broadly as identifying and linking one patient's data within and across health systems to obtain a comprehensive view of that patient's healthcare record [3]. This article examines what the documentation supports about record ownership, matching inputs and outputs, deployment boundaries, and unresolved questions, then offers a hypothetical validation scenario that does not invent test results.
What the Identity-Matching Standard Requires
The HL7 identity-matching guide profiles the $IDI-match operation to account for constraints that apply when authorized, trusted parties match patients across organizations [1][2]. Conformant systems must support the Patient $match operation, or an equivalent if non-FHIR systems are deployed in addition to or instead of FHIR, as further constrained by the $IDI-match profile [1][2]. The guide states that the best practices and suggested weights are based on the team's original research, which included cited sources, FAST webinar feedback, and FAST subject matter expert input [1][2]. This means the weighting guidance is a documented recommendation from a specific standards effort, not a universal accuracy guarantee.
The guide distinguishes patient matching from person matching, noting that matching requirements also apply to person matching that may need to occur for non-patient users, though some sections refer to the capability more narrowly as Patient Match [1][2]. The documentation also addresses populations with unstable contact information, stating that because many patients experiencing homelessness or housing instability have cell phones and email addresses used in identity verification and multi-factor authentication, those identifiers should be used for that population to enhance healthcare experience and communication with providers [1][2]. This is a specific recommendation about identifier selection, not a claim about match accuracy for that population.
Record Ownership and Partitioning Decisions
The identity-matching guide addresses record ownership through the concept of partitioning records into identities. A match implementation should partition its records into identities in real time as they arrive, which the guide says enables manual stewardship, improves matching quality, and improves searching quality [1][2]. The guide also states that a match implementation should enable manual stewardship of the partitioning based on identity [1][2]. This positions human review as a documented component of the identity resolution workflow rather than an optional afterthought.
The guide describes a scenario in which a doctor reviews the set of records that constitute an identity. It identifies three possible explanations for a problem: the patient has a diagnosis unknown to the doctor, the health information exchange has another patient's record mixed into the identity of the patient of interest, or the clinician is looking at the wrong patient's information [1][2]. The guide notes that in all three cases the patient's care might be improved if the doctor reviews the records, and that if the problem is the second case, both the care of this patient and perhaps others might be improved if the doctor could contribute to how records are partitioned [1][2]. This framing treats record ownership as a stewardship responsibility shared between the matching system and clinical users.
A Hypothetical Validation Scenario
A healthcare analytics team could design a validation scenario without claiming independent test results. The team would assemble a reference set of record pairs with known match status, then run both a probabilistic algorithm and a referential algorithm against the same demographic data. The probabilistic algorithm would compare fields such as name, birth date, phone number, and address using weights, while the referential algorithm would consult an external reference dataset to resolve identities. The team would record the number of declared matches, nonmatches, and pairs outside any blocking scheme, then manually review a sample to establish a gold-standard match status.
This scenario mirrors the documented approach in a study that evaluated real-world referential and probabilistic patient matching using approximately 47 million HIE records and 15,000 reference record pairs within blocking schemes, where 7,647 pairs were declared matches and 7,353 nonmatches [5]. The study noted that reference data quality, including completeness and accuracy, influences referential matching performance, and that results from Indiana health systems may not generalize to populations with different coverage for children or people experiencing homelessness [5]. A validation scenario should therefore report denominator, inclusion rules, and source dates rather than a single accuracy figure.
Matching Inputs, Outputs, and Verification Boundaries
The guide recommends including all known required and optional patient matching attributes in a match request, such as USCDI Patient Demographics, and indicates that the table provides examples of n-tuple or n-combination attributes under the category of combinatorics and levels of verification for consideration in different use cases [1][2]. The documentation also states that it is helpful to know the date that attributes were verified, particularly for address and mobile number since those attributes change over time, and that future versions will likely include a grammar for indicating verification date in match transactions [1][2]. This identifies verification date as a known gap in the current version.
For outputs, the guide specifies that linkage between records should be indicated by the Patient.link field, records should be ordered first by identity and then by score versus the input, and identities should be ordered by score versus the input [1][2]. If a match implementation supports creating a Golden Record to summarize the identity, the match output should contain that record as well, and the Golden Record may have an opinion on the patient's current address and consolidate demographics distributed across records [1][2]. The guide also states that a match output should reveal a presence or lack of manual stewardship, currently supportable via extensions [1][2]. These are documented output expectations, not performance benchmarks.
Deployment Boundaries and Unresolved Questions
The identity-matching guide identifies several capabilities as future work rather than current requirements. An API from USPS may be helpful in verifying individual street addresses in future versions, while National Provider Identifier records can currently be used to verify provider names, addresses, and telephone numbers [1][2]. The guide also states that future versions will likely include a grammar for indicating verification date in match transactions as well as the evidence that may be used to verify individual attributes [1][2]. These statements establish clear boundaries between what the current implementation guide supports and what remains under development.
The guide notes that transmitting identity attributes to third parties with whom sharing personally identifiable information is permitted can occur within an OpenID Connect user profile, another user information request to an Identity Provider, or the resultant assertion or claim, or within an HL7 B2B with User Authorization Extension Object [1][2]. The documentation does not specify which authorization mechanisms are required in every deployment. The guide also indicates that the team will consider input submitted via Jira tickets for exception handling guidance in the next version [1][2]. This leaves exception handling as an explicitly unresolved area in the current version.
Sources & references
- Patient Matching - Interoperable Digital Identity and Patient Matching v2.0.0 · hl7.org
- Patient Matching - Interoperable Digital Identity and Patient Matching v2.0.0 · hl7.org
- Patient Identity and Patient Record Matching - ONC - Office of the National Coordinator for Health Information Technology · healthit.gov
- Safer-Guide-6.-Patient-Identification-Final.pdf · healthit.gov
- Evaluation of real-world referential and probabilistic patient matching to advance patient identification strategy - PMC · pmc.ncbi.nlm.nih.gov
- Enhancing Patient Matching in Support of Operational Health Information Exchange - Final Report · digital.ahrq.gov
- Patient Identification · healthit.gov
- Data Management for Health Parity | Office of eHealth Innovation · oehi.colorado.gov