Patient-name matching is the subroutine that most US MPI deployments lean on more than they admit. Strong identifiers (SSN, MRN, government ID) are absent or unreliable for a meaningful fraction of records. Date of birth and address narrow the candidate set, but the name comparison is the feature that often decides whether two records resolve to one person. The choice of string-similarity metric inside that comparison matters more than vendor marketing suggests, and the two algorithms most often compared (Jaro-Winkler and Soundex) have different strengths against US patient data.
This piece walks through the comparison for US practices. The wider buyer-side context is in master patient index in US healthcare: a 2026 buyer's guide; the broader probabilistic-library question is in top 7 probabilistic patient-matching libraries for FHIR stacks. For the FHIR knowledge collection, the rest of the series adds context.
What Each Algorithm Actually Does
Jaro-Winkler is a string-similarity score that compares two strings character by character, with a bonus for common prefixes. It produces a score between 0 (no similarity) and 1 (identical). For patient name matching, it handles minor typos, transpositions, and short edits well.
Soundex is a phonetic algorithm that reduces a string to a code based on its sound. Two names that sound similar produce the same Soundex code regardless of spelling variation. It handles cases where the spelling diverges but the pronunciation does not.
The two algorithms answer different questions. Jaro-Winkler asks "how close are these strings as written." Soundex asks "do these strings sound the same when spoken." Both are useful; they catch different mismatches.
Where Each One Wins
Jaro-Winkler wins on typo patterns and short edits. A patient whose name is registered as "Christopher" in one source and "Christofer" in another produces a Jaro-Winkler score that is close enough for the matching engine to surface the records as a likely match. The Winkler bonus for shared prefixes specifically helps with surname matches where the first few characters are usually right.
Soundex wins on phonetic equivalence. Names that come from non-English-language origins, names with multiple legitimate Anglicized spellings (Stephens, Stevens, Stephenz), names registered by different staff who heard the same name differently: all of these are cases where the strings look different but the names sound the same. Soundex captures the equivalence.
For US patient data specifically, both kinds of mismatches are common, and neither algorithm alone captures them all.
How They Work Together in Practice
A working MPI usually uses both algorithms in combination. Jaro-Winkler catches the typo and edit patterns; Soundex (or its more accurate modern relatives like Metaphone or Double Metaphone) catches the phonetic patterns. The matching engine combines the two scores into a composite name-similarity feature that handles both kinds of variation.
The combination is not always literal: some MPI engines use Jaro-Winkler as the primary comparator and use Soundex only for blocking (the fast pre-filter that narrows the candidate set). Others use both as parallel features with separate weights. The exact arrangement varies; the principle that you need both is consistent.
What to Configure in Your MPI
Three configuration choices matter when picking between (or combining) these algorithms.
- The threshold above which a Jaro-Winkler score counts as a likely match. Common defaults are 0.85 to 0.90; below that the score is treated as evidence but not as a strong signal.
- The Soundex variant you use. The classic Soundex is fine for English-anchored names; Double Metaphone handles a broader range of name origins more accurately.
- The combination logic. Whether the composite score is a weighted sum, a max, or a separate feature each with its own weight.
These choices should be tuned against real data, not left at vendor defaults.
How to Decide
For greenfield MPIs that need to pick a starting point, Jaro-Winkler plus Double Metaphone is the modern default that handles US patient data well. For MPIs that already use Soundex, layering Jaro-Winkler on top usually improves matching without requiring a re-tuning of the whole engine. The deeper library question is in top 7 probabilistic patient-matching libraries for FHIR stacks, which goes into the implementations that combine both well.
Name-similarity choices are not glamorous, but they decide how the matching engine performs against real US patient data more than any other single configuration.
Sources
- Score-Based Persons Matching Algorithms Using Demographic Data (Jaro-Winkler vs Soundex empirical comparison, peer-reviewed) - PMC/NCBI
- Combining string and phonetic similarity matching (Jaro-Winkler plus phonetic, peer-reviewed) - PMC/NCBI
- Assessing Record Linkage Matches Using String Distance Measures - New Mexico Statistical Analysis Center PDF