Moving FHIR resources isn't the future of health data exchange

I don’t really think that FHIR will be the future of health data exchange.

That probably sounds like a crazy thing to say because FHIR is everywhere and basically every conversation about healthcare interoperability eventually ends up talking about FHIR APIs. I’m not saying that FHIR will disappear. I think something else is happening.

With how quickly LLMs, reasoning models and decision systems have improved over the last few years, I think we’re going to move from asking:

How do we move FHIR resources between systems?

to asking:

How do we move the meaning of health data between systems?

That distinction matters a lot.

FHIR can still be part of that world. It might even become more useful in it. I just don’t think the FHIR resource, or the FHIR server, has to remain at the center of the architecture.

The problem isn’t moving data. It’s knowing what the data means.

The more I explore what is actually done today in the industry, and the different ways organizations exchange information, the more I realize that what we need is not as much FHIR as we need a way to convey meaning.

Here is one example of many:

Medication status: active.

What does “active” actually mean?

• Does it mean the medication has been prescribed?
• Does it mean the patient is still taking it?
• Does it mean nobody has removed that medication from the patient record yet?

Those are three very different things. You can transport this JSON payload but the meaning is lost quickly.

FHIR profiles help with this. Profiles can constrain resources, make implementations more consistent and remove some ambiguity. But I don’t think they solve the whole problem.

Even two systems that both “support FHIR” can still disagree about what a field means and display it completely differently in their UI.

The next layer is semantic, not syntactic

What we do instead is semantics, and ontologies.

We have things like SNOMED CT for clinical concepts, LOINC for laboratory observations and different status systems that can describe where something is in its lifecycle. Those systems help answer:

What is this thing?

The layer we are lacking is meaning:

Why do we believe this is what it means?

That is where evidence becomes important.

It is not enough for a system to say “This field means X.” It should be able to say:

“This field means X because it came from this source, was created through this workflow, was mapped using this rule, was approved under these conditions, and here is the evidence that lets you inspect that decision.”

That changes interoperability quite fundamentally.

We stop exchanging only information. We start exchanging claims about information, together with evidence for those claims.

AI agents can do the heavy lifting. They shouldn’t become the truth.

This is also where I think AI becomes extremely useful.

Today, a huge amount of interoperability work is basically people trying to understand how one system represents something and then translating it into the representation expected by another system. An AI agent can investigate that: inspect documentation, compare examples, look at historical data, identify patterns, suggest that field A in system one corresponds to concept B in system two.

But I don’t think we should necessarily have an AI model reinterpret every data point every single time it moves.

Instead, I think AI will be more useful as a mapping discovery system. The AI investigates and proposes a mapping, and a human reviews the important parts. Once approved, that mapping becomes a deterministic rule and the runtime system simply executes it.

So the architecture becomes something like a large intelligent edge adapter.

The first time an unknown field appears, an agent investigates it and serves its conclusion to a human. Once that mapping is approved, it goes back into the system.

From that point onward, we don’t need to reason from scratch every time. The system already knows what to do.

This is particularly important because healthcare interoperability doesn’t only need flexibility. It also needs determinism.

Meaning also needs responsibility

Once we start exchanging semantic claims instead of just fields, another question appears:

Who is responsible for that meaning?

We will need ways to ensure data safety and integrity because in some cases a mistake in data exchange can harm a patient. Therefore we need liability on the information sent and this could be done by creating signed statements and the usage of cryptographic safety. But for critical information, we should be able to establish responsibility and provenance.

A signed statement could say:

• who produced the information;
• what semantic systems were used;
• what mapping or transformation was applied;
• what evidence supports the interpretation;
• where the raw source can be found;
• and whether the data has been altered since that statement was created.

That creates an audit trail. And I think auditability will become much more important than people currently assume.

Because once AI starts doing more interpretation, the important question is not only:

What did the model output?

It becomes:

How did we arrive here, and can we reproduce that decision?

Evidence and interpretation should stay separate

One idea I keep coming back to is that evidence and interpretation should be separate. That separation survives almost every version of this future.

You can disagree about whether interpretation happens with FHIR profiles, AI agents, vendor-certified mappings, national standards or runtime models. But the original evidence should still exist independently.

The system might say “We interpret this as an active prescription.” But the receiving organization should still be able to inspect the underlying event or source that caused that interpretation.

This matters because interpretations change.

Evidence shouldn’t.

I also think there will be a shared trust layer

We will see a strong shared trust layer. Everyone needs it, and multiple organizations are already working on it. Once you have provenance, signed statements, identity, consent and authorization, it naturally leads to some form of shared trust infrastructure.

Some of this already exists today. We have infrastructure and identity validation in place in many countries. I think this layer will expand.

And by 2030, I would not be surprised if some categories of highly reusable health information are treated almost like national reference data. Things like:

• identity;
• allergies;
• vaccinations;
• active prescriptions;
• major procedures.

Not everything, but most of the above are already done or in progress and it will grow naturally.

This actually changed how I think about this architecture. Originally, I was thinking mostly about a decentralized world where every hospital keeps its own data and semantic adapters make everything understandable.

I now think the more realistic future is probably tiered.

The realistic 2030 is probably a tiered system

I think there will be three layers.

1. A mandated national core

Some information is simply too important and too reusable to let every organization interpret it differently. Identity, allergies, vaccines, active prescriptions and perhaps major procedures could belong here.

Regulators or national operators may define very strict representations and reference services for these. There is very little reason for a hospital to invent its own definition of a vaccine record if an entire country can agree on one.

2. A vendor-certified middle

Then there is a large group of data where national standardization is possible, but where vendors are likely to do most of the implementation work.

Instead of every hospital mapping every field independently, a vendor could certify:

“Version 8.3 of our system exports this concept in this way.”

Every hospital running that version benefits from the same semantic work. This is much more scalable than repeating the same mapping exercise thousands of times.

3. An AI-discovered long tail

And finally, there will always be the messy part.

• Local fields.
• Legacy databases.
• Specialty workflows.
• Hospital-specific codes.
• Old systems nobody documented correctly.
• Data that simply wasn’t important enough to standardize nationally.

That is where I think AI agents become extremely valuable.

The funny thing is that this means the architecture I initially imagined may describe only the long tail of interoperability rather than the whole system. And I think that is actually a stronger argument.

Because we don’t need AI to solve the easy 80% if regulation, vendors and standards can solve it more reliably.

AI is most valuable where the deterministic world stops.

In that world, I don’t really care what format you store the data in

Once meaning, evidence and trust become portable, the storage format becomes much less important.

Store it as FHIR. Store it as openEHR. Store it in an HL7-based system. Store it internally in your own representation. I don’t really care.

Because the important boundary is no longer:

“Can your database look like my database?”

It is:

“Can you reliably tell me what this information means, where it came from, and why I should trust that interpretation?”

At that point, FHIR becomes one interface into the semantic system rather than the semantic system itself. It might still be one of the most useful interfaces. It just doesn’t need to be the center of everything.

Interoperability becomes less about exchange

There is another interesting consequence.

If we have shared trust infrastructure, national reference information and systems that can interpret evidence dynamically, we may actually exchange less data.

Instead of pushing complete copies of patient records everywhere, systems could query the source when they need something. Or they could ask a specific question. Or an agent could retrieve the relevant evidence, interpret it for a particular clinical decision, and return the answer together with its provenance.

In other words, sometimes the question travels instead of the data.

That’s the big change I’m looking for in the way interoperability is usually discussed today.