Healthcare master patient index (MPI) software research should start with the matching workflow and the evidence that supports it, not with a feature count. An MPI assigns an enterprise identifier to person records and links demographic data arriving through HL7 admission, discharge, and transfer messages [1][6]. The core research question is whether a given matching configuration detects true duplicate or related records at an acceptable error rate for a specific patient population. Vendor materials describe deterministic, probabilistic, and referential matching capabilities [5][7]. Peer-reviewed validation reports show that baseline configurations can perform far worse than optimized ones when evaluated against labeled data [3][4]. An evaluation plan therefore needs documented matching logic, a reproducible test set with known links and nonlinks, and clear thresholds for automatic versus manual review [3][4][7].

What the MPI Workflow Actually Exchanges

The Indian Health Service technical manual describes an MPI interface that transfers patient data through HL7 messages and assigns an Integration Control Number, also called an Enterprise Unique Identifier, during initialization [1]. Specific message types include A28 for adding a patient, A08 for updating cross-referenced fields, A01 for check-in and check-out events, and A03 for admissions and discharges. The technical manual also lists an HL7 M05 Location Master File Message from the MPI, through which RPMS receives treating facility information when a patient is seen at a new facility or two patient records are linked in the MPI. Beginning in patch 2, the treating facility list is deprecated and M05 messages are not processed, although the central MPI may still send messages to RPMS [1]. This message set defines a narrow operational workflow: registration edits and ADT events push demographic changes, while treating facility information may be returned to RPMS depending on the patch level [1].

The user manual adds that the interface includes menu options for configuring application settings and creating A28, A08, A01, and A03 messages [6]. Outbound queue details show ADT A28 messages with MSH, EVN, PID, and ZPD segments, which identifies the demographic payload but not the matching decision itself [6]. A researcher comparing products should trace which events trigger matching, which fields are sent, and how linked-record decisions return to the source system. Documentation for InterSystems Patient Index instead describes bulk import APIs, FHIR and HL7 v2 support, and IHE profiles including PIX, PDQ, and XCPD [7]. These are different integration surfaces, and neither document alone establishes matching accuracy.

Matching Methods and the Gap Between Claim and Evidence

IHS documentation states that MPI software uses probabilistic and other matching algorithm protocols to evaluate identities [1]. InterSystems materials describe a matching engine using deterministic and probabilistic logic with a referential data set, and mention LexisNexis LexID integration for referential matching in the United States [7]. The InterSystems EMPI product page describes an automated identity resolution solution and asserts that 35 percent of denied medical claims stem from inaccurate patient identification [5]. That statistic is a vendor claim without a cited methodology in the retrieved page, so it should be treated as promotional context rather than independent evidence.

A stronger evidence base appears in a peer-reviewed JMIR Formative Research study of SantéMPI [3][4]. The researchers note that automated matching algorithms guide manual record linkage and must be adapted to the population being served [3][4]. They used a Bayesian optimization tool to tune matching configurations and evaluated performance on held-out synthetic data sets. In one reported result, an optimized configuration reached 100 percent sensitivity with 95.9 percent specificity for possible linkages, compared with a baseline sensitivity of 90.2 percent and specificity of 100 percent [4]. The same paper reports that baseline configurations failed to detect definite matches in some data sets [4]. These numbers illustrate that configuration tuning, not the software brand alone, drives measured performance.

How to Build a Local Matching Evaluation Set

Begin by exporting a sample of previously merged duplicate records and confirmed distinct pairs from the EHR. Assign each pair a ground-truth label of match or nonmatch before running any algorithm, and remove pairs already used during configuration tuning to avoid inflated results [4]. Run the candidate MPI in a test environment with the same demographic fields that production ADT messages will provide [1].

Score each pair and group outputs into definite links and possible links, matching the reporting structure used in the JMIR validation study [4]. Calculate sensitivity, specificity, and positive predictive value for both groups. Record the manual review workload implied by the possible-link count. Finally, repeat the process after any threshold change or source data cleanup, because results apply only to the configuration and data snapshot tested.

Designing a Reproducible Evaluation

The JMIR study offers a template for MPI software research: hold out known matches and nonmatches, run the algorithm, and report sensitivity, specificity, and positive predictive value [3][4]. The authors made their configuration optimization tool, data, and data set generator freely available, and confirmed that all data used were artificially synthesized using FEBRL or distributed with the FEBRL package [4]. Because the evaluation used synthetic records, transfer to a production population with real name variation, missing fields, and cross-facility formats remains an open question. A research report should state that limitation explicitly rather than treating the published percentages as universal benchmarks.

A practical evaluation should also separate definite linkages that can be automated from possible linkages that need manual review. The JMIR paper reports separate performance tables for these two categories [4]. This distinction matters for staffing and workflow design. If a configuration aims for high sensitivity on possible matches, reviewers will spend time on more candidate pairs. If it optimizes for precision, missed links may persist. Neither the IHS manuals nor the InterSystems fact sheet provides comparable public threshold data [1][6][7]. Researchers should therefore request documentation of the matching scores, threshold settings, and evaluation sets used during validation.

Integration Constraints That Shape Matching Quality

Matching cannot improve data that never reaches the MPI correctly. InterSystems describes a Data Quality Framework that analyzes source data cleanliness and supports routines to clean data before it enters the EMPI [7]. The IHS manual frames the MPI as the primary repository for enterprise-wide patient data once interface implementation is complete, with patient data accessible through the National Health Information Network and a Personal Health Record application [6]. These are architectural statements, not evidence that source data in a different organization will meet the same assumptions.

The SAFER Patient Identification guide ties record linkage failures to patient harm risk and recommends using an EMPI to identify patients with existing records [2]. It also notes that materials printed from the EHR should include multiple identifiers and an electronic verification method, such as a barcode, and that medical record numbers should incorporate a check digit [2]. These practices sit outside the MPI itself but affect duplicate creation rates. An MPI evaluation should therefore document upstream registration controls, source system message completeness, and the process for correcting overlaid records after the fact.

A Specific Documentation-Based Evaluation Example

Suppose a regional health information organization must compare an InterSystems EMPI deployment against an internally hosted open-source MPI with HL7 v2 registration feeds. The team could start with the message contract: which A28 and A08 segments each product accepts, and whether A01 and A03 events trigger re-matching or only demographic updates. The IHS documentation demonstrates one possible A28 segment sequence, but it does not establish that another vendor handles the same fields with the same semantics [1][6]. InterSystems documentation indicates support for HL7 v2 and FHIR, but field-level behavior would need confirmation from implementation guides [7].

The next step would be a labeled test set built from historical duplicate records already corrected by staff. The team could measure how often each configuration places known duplicate pairs above the automatic-link threshold and how many known nonmatches appear as possible links. This approach mirrors the JMIR evaluation design without reproducing its synthetic data [4]. The result would be a local sensitivity and review-burden estimate. It would not prove which product is safer across the market, because no retrieved source contains head-to-head measured accuracy between these specific systems.

Evidence Limits and What Remains Unknown

The retrieved sources leave several questions unresolved. No source provides comparative false-positive or false-negative rates for InterSystems EMPI against an open-source alternative in the same patient population. No source documents pricing, implementation time, or support performance. The IHS manuals describe a specific RPMS environment and a June 2026 patch level, which may not generalize to commercial EHRs or non-IHS facilities [1][6]. The JMIR study is limited to synthetic data sets and one MPI implementation, so its sensitivity and specificity figures cannot be cited as product-wide accuracy claims for other software [3][4].

ONC defines patient matching as linking one patient's data within and across health systems using demographic fields such as name, birth date, phone number, and address [8]. This definition helps frame research scope, but it does not set an accuracy target. The SAFER guide explicitly states that it is not a legal compliance tool and does not guarantee HIPAA compliance [2]. Researchers should treat any legal or regulatory interpretation as outside the evidence base. The defensible conclusion is narrower: MPI selection depends on documented event coverage, integration surfaces, and locally measured matching performance.