Home health agencies sit at the awkward edge of patient matching. The patient is at home, the data sources are fragmented across the hospital that discharged the patient, the physician that referred the visit, the pharmacy that filled the medications, and the lab that ran the panel last week. The fuzzy matching approach, all-classic in MPI, walks each pair of records and decides how similar they are. The reference-based approach uses an external identity graph, usually built on consumer demographic data, to anchor the match. The picking question for home health is which model holds up better when the data is this messy.
This walks through how fuzzy matching and reference-based MPI compare for home health work, where each one fits, and how to think about combining them. For FHIR background reading covering the wider context, the broader catalog is the place to start.
For the upstream picture, the complete guide to FHIR master patient index for payers in 2026 is the reference.
What Fuzzy Matching Brings
Fuzzy matching, the classic probabilistic approach used by NextGate, OpenEMPI, and most open-source matchers, evaluates pairs of records on the demographic fields the records contain. The honest strength is that the matcher works with what you have, no external dependency. The honest weakness is that the matcher only knows what the records know, and home health data is famously incomplete.
A home health agency that runs fuzzy matching alone tends to learn fast that an address typo plus a slightly different first name can prevent a clean match, even when a human reviewing the case would call it the same patient in five seconds.
What Reference-Based MPI Brings
Reference-based MPI, the approach used by Verato and a handful of other vendors, anchors the match against an external identity graph built from consumer demographic data. The matcher resolves identity by asking the graph, not just by comparing the two records in hand. The honest strength is that the graph knows things the records do not, including past addresses and known name variants. The honest weakness is the per-call cost and the external dependency.
A home health agency that runs reference-based MPI tends to see fewer false negatives on the messy cases, at the price of a recurring per-call fee.
Where Each Model Fits for Home Health
Three settings make the picking question concrete:
- Initial admission match. Reference-based often wins. The first record from a new patient is the hardest to match, and the external graph helps most here.
- Ongoing visit match. Fuzzy usually wins. By the third visit, the agency has enough internal data that the external graph is no longer adding much.
- Cross-source reconciliation. The honest answer is both. The agency wants fuzzy matching on the internal sources plus reference-based matching as a fallback for the records that fuzzy cannot resolve cleanly.
A home health agency rarely lands on a single model in the end. The common pattern is a fuzzy matcher for the steady state plus a reference-based fallback for the hard cases.
How to Run a Real Comparison
Pick a hundred real admission records that fuzzy matching could not resolve cleanly last year. Send them through a reference-based matcher and measure how many resolve correctly versus how many produce a wrong match. The honest answer often surprises both procurement and clinical leadership.
What This Tends to Mean for Procurement
A home health agency that picks one model and pretends the other does not exist usually pays for it in manual reconciliation. The cleaner approach is to plan a stack that uses each where it fits. That makes the procurement conversation easier and the matching layer honest.
For closely related discussions, the 5 MPI engines that handle veteran identifier edge cases and top 5 FHIR MPI products for state Medicaid programs in 2026 are the natural next reads for adjacent identity work.
Sources
- MPI vs EMPI vs UMPI (referential matching explainer) - Verato blog
- [Patient Demographics Match [ITI-119] v3.2.0](https://profiles.ihe.net/ITI/PDQm/ITI-119.html) - IHE Profiles
- $Match OperationDefinition (probabilistic similarity confidence scores) - IHE PDQm build