Why Two FHIR Feeds Don't Match
Date Published
Nov 17, 2026
Written by
Consolidate Health
Time to Read
4 min

FHIR is a standard, and certified electronic health records are required to expose it. Both of those things are true, and together they suggest something that is not true: two FHIR responses describing the same clinical event will look alike.
They will both be valid and both will parse; however they will not completely match, and the places they diverge are consistent enough to be worth knowing before you plan around them.
What the Standard Actually Fixes
FHIR specifies resources and their structure. A MedicationRequest has defined fields and defined types. A Condition does too. An Observation carries a value and a code in agreed places.
What the specification deliberately does not do is dictate how any particular organization models its own clinical workflow, which vocabularies it uses internally, or which optional elements it populates. It constrains the container. Most of the interesting variance lives in what each system decides to put in it.
That latitude is intentional. Healthcare organizations document care differently because they deliver it differently, and a specification rigid enough to eliminate the variance would be one almost nobody could implement.
Where It Diverges in Practice
Vocabularies. The same medication can arrive coded against different terminologies depending on the source, sometimes with a local code and no standard mapping at all. Two systems can describe the identical prescription in codes that share no values. If your logic keys on codes, it needs a translation layer before it needs anything else.
Where a fact lives. One system populates a dedicated field. Another leaves that field empty and carries the same fact in an extension. A third does not model it discretely, so it exists only as free text inside a note. Your code finds the value, finds nothing, or finds nothing but the information is present, but somewhere it was not looking.
Granularity. A single real-world event can appear as one resource or as several linked ones. A medication changed in dose might be one record with a history, or two records that look like two separate prescriptions unless you reconcile them. Naive handling produces a medication list with phantom duplicates.
Completeness of the exposed surface. Certification sets a floor, not a ceiling, and implementations differ in what they make available beyond it. Some sources return rich attachment metadata; others return the attachment with almost nothing describing it.
Behavior rather than content. Pagination, search parameter support, date filtering, how errors are signalled, rate limits - none of this is the clinical data, and all of it determines whether your retrieval code works against a given source.
This Is Not a Bug Anyone Is Going to Fix
Worth being clear, because teams sometimes plan as though the variance is temporary.
It is not a transitional problem waiting on better compliance. It is what happens when a specification is flexible enough to be implemented by organizations that genuinely work differently from each other. Future versions will narrow some of it. They will not remove it, because removing it would mean telling health systems how to model care.
So the mapping layer is not scaffolding you take down once everyone catches up. It is a permanent component of any system that reads from more than one source.
What That Means If You Build It Yourself
The first integration teaches you less than you would hope. It works, it produces clean data, and it gives you a mental model that the second integration breaks.
By the third or fourth source, the shape of the real work becomes clear. You are not writing FHIR clients. Instead you are maintaining a growing set of source-specific mappings, each with its own exceptions, each needing revision when an organization upgrades its system or changes how it models something. The code is not necessarily hard, but the ongoing reconciliation is, and it scales with the number of sources rather than flattening out.
This is also the work that is least visible from outside. Nobody demos their mapping layer.
What We Hand You Instead
One schema, consistent across sources. Medications arrive as a medication list, with start dates, stop dates, and dose changes in sequence, regardless of which system they came from or how that system chose to model them. Conditions, labs, allergies, and encounters likewise.
Where a source leaves something out of the structured data but records it in a note, the note comes through alongside, so the information is available rather than silently absent.
The variance still exists. We absorb it on our side, once, rather than every consumer of the data absorbing it separately.
How to Check Anyone's Claim
Normalization is easy to assert and easy to verify.
Ask for the same category of data, a medication list is the clearest, for test patients from two different source systems. Then compare the two responses field by field. If the structure, the coding, and the granularity match, the normalization is really happening. If you can tell which EHR each response came from, you are being handed a pass-through with a FHIR label on it, and the mapping work is still your responsibility.
It takes about an hour and it tells you more than any amount of documentation.
If you'd like to learn more about Consolidate Health's API and how we retrieve a complete normalized longitudinal record through patient-authorized release, book an intro call today.
If you prefer to try the product with zero commitment, you can request sandbox access.

