Patient matching algorithms have been the quiet center of US healthcare identity work for two decades. The choice of algorithm shapes everything downstream: false-positive rate, false-negative rate, audit complexity, and the workload that lands on the human resolvers who clean up edge cases. These seven algorithms are the ones US healthcare IT teams encounter most often in 2026, in commercial products and in open-source stacks.
For the broader buying frame, see the cornerstone Complete US FHIR Master Patient Index Buyer's Guide for 2026. For more FHIR explainers for US teams, the rest of the desk follows on.
Deterministic Matching
Deterministic matching compares fields exactly. Same SSN, same date of birth, same first name, same last name. The strength is auditability. The match decision is obvious, and a human resolver can reproduce it from the inputs. The weakness is recall. Any data noise, including a typo or a hyphenated name change, drops the match rate. For US health systems with high-quality input data, deterministic matching is a credible primary engine.
Probabilistic Matching With Fellegi-Sunter
The Fellegi-Sunter probabilistic model is the foundation under most US enterprise MPIs. Each field comparison produces a likelihood ratio, the ratios combine into a composite score, and threshold cutoffs separate matches from non-matches. The strength is recall against noisy data. The weakness is tuning effort, which is non-trivial for US patient populations with large demographic shifts.
Hybrid Deterministic-Probabilistic Matching
Most US production MPIs run a hybrid algorithm. Deterministic matching handles the high-confidence cases cheaply, probabilistic matching handles the rest. The combined accuracy is generally higher than either approach alone, and the operational cost is lower than running probabilistic matching on every record. For US health systems at any scale, the hybrid pattern is the standard.
Referential Matching
Referential matching is the pattern Verato and a handful of other vendors have built around. Each candidate record is matched against a reference dataset that already contains identity information for a large slice of the US population. The strength is high accuracy without per-system tuning. The weakness is dependency on the reference dataset's coverage and freshness, which is the vendor's competitive moat.
Machine Learning Match Scoring
A handful of US MPI vendors and large health systems have moved to machine learning match scoring on top of a traditional probabilistic backbone. The model learns from historical resolver decisions and improves the match cutoff over time. The strength is adaptability to local patient patterns. The weakness is explainability, which US audit reviewers ask about more often than vendor marketing acknowledges.
Phonetic and Edit-Distance Comparators
Phonetic comparators like Soundex and Metaphone, paired with edit-distance comparators like Levenshtein and Jaro-Winkler, handle the name-variant problem that is the single biggest source of false negatives in US patient matching. These are not standalone algorithms in production; they are components inside a deterministic or probabilistic engine. Every credible US MPI uses some combination.
Address Standardization Comparators
Address handling is the second biggest source of US matching errors after name variants. Address standardization comparators, often backed by USPS address verification data, normalize the input before comparison. Inside a US MPI, address handling is one of the highest-leverage tuning surfaces. Get this right, and false-negative rates drop noticeably.
Picking Algorithms for a US Healthcare IT Stack
The realistic answer for most US healthcare IT teams is not picking one algorithm. It is picking a product whose algorithm portfolio fits the patient population and the data quality. Hybrid deterministic-probabilistic is the standard, referential matching wins where the reference dataset fits, and machine learning scoring is moving into the mainstream for large US health systems.
The Top 5 FHIR-native MPI products for US health systems in 2026 covers the product side of this decision. The right algorithm stack for US healthcare IT is the one the team can keep tuned and defensible through three years of patient data drift.
Sources
- Patient matching evaluation study - Journal article PMC, evaluating real-world referential and probabilistic patient matching, 2022
- Patient Matching Aggregation and Linking Final Report (foundational US reference) - PDF report, HHS ASPE / PMAL project, 2019
- Patient Matching profile reference in Identity Matching IG - IG section, HL7, 2025 v2.0.0