Duplicate patient records are the steady-state problem an MPI exists to solve. Every healthcare organization has them, every new EHR rollout produces more, and every reconciliation effort surfaces a new round of edge cases. The four algorithmic approaches below cover most of the production duplicate-resolution work in 2026, and the right MPI for a given deployment usually combines two or three of them. For broader background, see additional FHIR API notes.
The master patient index reference guide covers the architectural picture in which these algorithms operate.
The 4 Algorithmic Approaches to Know
- Deterministic rule-based matching. The classic approach: define rules over identifiers and demographic fields, declare a match when the rules agree. Fast, cheap, easy to audit, and the right starting point for cleaning the most obvious duplicates.
- Probabilistic Fellegi-Sunter matching. Score candidate record pairs against a weighted attribute model and declare a match when the score exceeds a threshold. Catches matches that deterministic rules miss but requires careful tuning of weights and thresholds.
- Phonetic and approximate string matching. Match on Soundex, Metaphone, or Jaro-Winkler distance over name fields to catch transliteration differences and typos. Usually layered onto deterministic or probabilistic matching rather than used alone.
- Referential matching against an external reference index. Compare both candidate records against a third-party reference data set; if both match the same reference entity, declare a match. Catches address-change cases the other three approaches miss.
How They Combine in Practice
Real MPI deployments rarely use one algorithm alone. The typical layering looks like:
- Deterministic on hard identifiers first. If two records share an SSN and a date of birth, they are the same person; no further analysis needed.
- Probabilistic on demographic clusters next. Records that miss the deterministic pass enter a probabilistic scoring step over demographic attributes.
- Phonetic and approximate matching on uncertain fields. Name fields and address fields get phonetic matching as part of the probabilistic weighting.
- Referential matching for the long tail. The cases that fall below the probabilistic threshold but might still match get checked against a referential index.
The audit trail for each decision has to record which algorithm produced the match and which attributes carried the weight. Without that, compliance teams cannot defend the decisions.
Where Each Approach Falls Short
Deterministic rules miss every case where the underlying data is dirty. Probabilistic matching requires per-deployment tuning that few organizations do well. Phonetic matching produces false positives on common name combinations. Referential matching adds vendor dependency and licensing cost.
The right approach for a single hospital running a single EHR is deterministic-plus-light-probabilistic. The right approach for a multi-state health information exchange is probabilistic-plus-referential. The right approach for a clinical research network with strict false-positive constraints is deterministic-only with manual review for ambiguous cases.
The probabilistic vs referential patient matching comparison covers the two approaches that dominate the long tail of duplicate-resolution work.
How Teams Should Pick
Selection turns on the data quality of the source systems, the volume of duplicates the deployment produces, and the tolerance for vendor dependencies. Most production MPIs combine three of these four approaches, and the choice is really about which third approach gets added to the deterministic-and-probabilistic baseline. For HIE-scale work, the top patient matching tools for cross-state health exchanges walkthrough covers the products that handle these algorithms at scale. The right choice tends to be visible in retrospect by what the team stopped thinking about, not by what they advocated for during selection.
Sources
- Optimizing Patient Record Linkage with ML - PMC, JMIR, 2023
- Real-world referential vs probabilistic - JAMIA, Oxford Academic, 2022
- Identity Matching IG - IG, HL7, 2024