Reconciling subject records across investigator sites is where matching algorithms earn their keep. The data is dirtier than the demos suggest, the rules of thumb that work in hospital MPI work less well in trials, and the cost of a bad match is high. Seven algorithms (and the tools that ship them) are worth knowing in 2026.
For framing, the complete guide to FHIR master patient index for clinical research in 2026 is the right primer.
1. Fellegi-Sunter Probabilistic Matching
The Fellegi-Sunter framework remains the textbook probabilistic matching algorithm and the basis of most serious tools. Field-by-field weights tune to the data, and the resulting score captures uncertainty better than rule-based alternatives.
Tools that implement Fellegi-Sunter well (MDMbox, IBM Patient Matching, OpenEMPI) outperform tools that hand-roll their own probabilistic approaches.
2. Jaro-Winkler String Similarity
Jaro-Winkler is the standard string similarity measure for name matching. It handles common typos, transpositions, and prefix matches better than simpler measures, and serves as a feature in larger matching frameworks rather than a complete algorithm on its own.
Most MPI tools use Jaro-Winkler under the hood for the name-comparison step; the differentiation lives in how the rest of the algorithm consumes that signal.
3. Soundex and Double Metaphone
Phonetic matching covers the case where names are misspelled in ways that string similarity misses. Soundex is the classic; Double Metaphone is the better choice for international populations because it handles non-Anglo name patterns more honestly.
For trials with diverse subject populations, Double Metaphone is worth a serious look as a feature in the larger matching pipeline.
4. Referential Matching
Verato's referential approach compares each record to a curated reference of known identities rather than to other records directly. The result is a strong signal in cases where in-trial records are sparse or inconsistent.
The trade-off is geographic coverage. Referential matching works well in the U.S. and is thinner elsewhere.
5. Address Standardization Plus Matching
Address-based matching is unreliable without a standardization step. Tools that normalize addresses before comparison (NextGate, IBM, MDMbox) handle real-world data better than tools that compare raw strings.
For trials with subjects whose primary identifier is residential rather than personal (some pediatric studies, some long-term-care studies), address standardization is the differentiator.
6. Rule-Based Deterministic Matching
Deterministic matching is the right algorithm when identifiers are clean and consistent: government IDs, study IDs, or other invariant tokens. The strength is auditability; the weakness is brittleness against any data quality issue.
For randomization-driven trial workflows, deterministic matching on the study identifier is usually the right path. For everything else, deterministic-only is too brittle.
7. Hybrid Probabilistic + Reviewer-in-the-Loop
The strongest algorithms in production are not algorithms at all but workflows: probabilistic matching with explicit reviewer attention on borderline cases. The MPI surfaces likely matches above one threshold, sends borderline cases to a reviewer interface, and rejects cases below a lower threshold.
This pattern is what every serious tool implements; the differences are in how the reviewer interface behaves under load and how the audit trail captures reviewer decisions.
What Matters in Practice
Across all seven, no single algorithm wins in isolation. What matters is the pipeline: phonetic plus string similarity as features, probabilistic scoring as the decision layer, deterministic checks for short-circuit cases, and a reviewer-in-the-loop workflow for the borderline cases. Tools that implement this whole pipeline cleanly outperform tools that lean heavily on one algorithm.
For the deterministic-vs-probabilistic angle, deterministic vs probabilistic subject matching for trial cohorts is the next read. For the collision-handling angle, 5 patient matching tools that handle investigator-site identifier collisions covers a common edge case. And related FHIR explainers on the homepage point to the rest.