Cross-hospital patient matching is harder than single-hospital matching by an order of magnitude. The demographics are messier, the identifier overlap is incomplete, and the governance has to satisfy several boards instead of one. The algorithms below are the ones that hold up at network scale in 2026, with notes on where each one earns its place and where each one needs a careful operational design around it.
Anyone new to the broader topic can skim more FHIR implementation context first.
What Network-Scale Matching Has to Solve
A cross-hospital network sees the same patient several times across different source systems, each with its own data-entry conventions. Names get clipped, dates of birth get transposed, phone numbers go stale, and addresses change. The algorithm has to figure out which records refer to the same person without falsely merging records that belong to different people. Errors in both directions are clinically meaningful.
The matching algorithm is one input into a larger system that also includes data preparation, normalization, governance, and review workflows. An algorithm that performs well in isolation can still fail in production if the operational design around it is weak. The algorithms below all need that operational scaffolding to work well.
The Algorithms That Matter at Network Scale
- Fellegi-Sunter Probabilistic Matching. The classical probabilistic framework that underlies most commercial MPI products. Computes a score per candidate pair using weights derived from agreement and disagreement frequencies in the population. Strong recall when tuned against the actual data distribution. The deterministic vs probabilistic patient matching for FHIR systems walkthrough covers the trade-offs.
- Referential Matching. The model that compares incoming demographics against a national or regional reference dataset. Useful when source data is uneven and the team needs help filling in identity gaps the source systems do not capture. Common in commercial offerings that include reference data licensing.
- Hybrid Deterministic-Probabilistic. The pattern most modern MPIs use in 2026. Deterministic rules handle the obvious matches and non-matches cheaply, probabilistic scoring handles the gray zone, and review queues handle the cases that score in the uncertain middle. Strong fit for cross-hospital networks where data quality varies by source.
- Graph-Based Linkage. The model that treats patient records as nodes in a graph and links across nodes through shared identifiers, demographics, or relationships. Useful when the network includes payers, providers, and patient-facing services that capture different identifier sets. Increasingly common in 2026 deployments.
- Machine Learning Augmented Matching. The newer pattern that uses a learned model to score candidate pairs, often layered over a traditional engine. Adds capacity to capture patterns the rule-based engines miss, but introduces explainability challenges that the operational design has to address. Most successful when the model output feeds a review queue rather than acting autonomously.
How to Pick the Right Algorithm
The first cut is data quality. Networks with clean data across most sources can lean deterministic and use probabilistic as a backstop. Networks with uneven data across many sources usually need a probabilistic or referential engine as the primary, with deterministic as a fast-path filter.
The second cut is governance. Algorithms that produce explainable decisions fit networks where compliance reviewers need to see the rationale for any specific match. Algorithms that are more opaque need additional governance scaffolding around them, including model documentation, review workflows, and audit trails.
The top 5 master patient index tools for hospital networks in 2026 shortlist covers commercial products that implement each pattern. The FHIR Master Patient Index overview is a useful read for teams thinking about the broader architecture.
The honest pattern across cross-hospital implementations in 2026 is that the algorithm is rarely the deciding factor for success. The operational design that surrounds the algorithm, including review queues, governance, and ongoing tuning, is what makes the difference between a clean implementation and a brittle one.
Sources
- Academic paper, 2024 - Linking EHRs for MS Research: Deterministic, Probabilistic, and ML Linkage Methods
- Academic paper, JAMIA 2023 - Optimizing Patient Record Linkage in MPI Using Machine Learning
- PDF, ONC, current - Perspectives on Patient Matching white paper