FHIR、HL7 與 AI：探索醫療資訊互通性的下一層

I've recently been spending some time exploring HL7, FHIR and the broader healthcare interoperability landscape—particularly how standards, implementation approaches and validation come together in practice.

As I organized some of those thoughts, another connection became increasingly interesting:

What happens when healthcare interoperability meets AI?

At first, these can seem like two separate conversations.

FHIR and HL7 help establish structure, meaning and consistency in how healthcare information is exchanged.

AI brings capabilities around language, reasoning, knowledge discovery and, increasingly, action.

The more I explored the two, the more interesting their intersection became.

Interoperability Is More Than Moving Data

Healthcare information exists across many different systems:

• Electronic health records
• Laboratory and pharmacy systems
• Clinical applications
• Registries
• Analytics platforms
• Public health systems

The challenge isn't simply moving data between them.

The receiving system also needs to understand what that information represents.

HL7 has developed healthcare interoperability standards for decades. FHIR (Fast Healthcare Interoperability Resources) is one of the standards developed by HL7, designed around modern web technologies and standardized approaches for representing and exchanging healthcare information.

FHIR organizes information into reusable resources such as Patient, Observation, Condition, Encounter and Organization.

Profiles, implementation guides and standardized terminology can then further define how those resources should be represented and exchanged within a particular healthcare ecosystem.

At a simplified level:

Healthcare Information
        ↓
HL7 / FHIR
        ↓
Resources + Profiles + Terminology
        ↓
Consistent Information Exchange
        ↓
Interoperability

But interoperability becomes particularly interesting when we consider how AI might use that information.

Why Interoperability Matters to AI

Generative AI is particularly good at working with language and unstructured information.

Healthcare data, however, isn't simply a collection of text and documents.

Meaning and relationships matter.

An Observation isn't an Encounter.

A Patient isn't a Practitioner.

A clinical code isn't simply another text value.

This is where standardized healthcare information could provide an important foundation for AI.

Instead of giving an AI system disconnected information from multiple proprietary formats, interoperability can provide more consistent structure, terminology, relationships and context.

FHIR alone doesn't make an AI system trustworthy.

But interoperable, well-governed and semantically meaningful data can provide a much stronger foundation on which useful AI capabilities can be built.

FHIR → AI: Where Are the Opportunities?

Several areas stand out to me.

Grounding AI in Healthcare Information

We've already seen how Retrieval-Augmented Generation (RAG) can ground AI responses using trusted information.

FHIR potentially adds another dimension.

Instead of retrieving only documents, an AI solution could potentially combine structured healthcare information with approved knowledge sources and relevant user context.

Structured FHIR Data
        +
Approved Knowledge
        +
Authorized Context
        ↓
Grounded AI Capability

This could create more contextual experiences than relying on a general-purpose model alone.

The quality of the model still matters—but so does the quality, structure and meaning of the information provided to it.

Natural Language Over Structured Health Data

Another interesting opportunity is making complex healthcare information easier to explore.

FHIR provides standardized ways to represent and exchange healthcare information, including modern API-based interactions.

AI could potentially provide a natural-language layer over some of that complexity.

Conceptually, a user could ask a question in natural language, while the system determines what authorized information is required, retrieves structured data and presents it in an understandable form.

This creates an interesting possibility:

Interoperability not only between systems, but also between people and complex health information.

Of course, what information can be accessed—and by whom—remains an important architectural and governance decision.

AI-Assisted Knowledge, Analysis and Workflows

FHIR-based information could also provide context for AI-assisted capabilities such as:

• Summarizing available information
• Supporting research and analysis
• Navigating complex healthcare knowledge
• Identifying information requiring review
• Assisting operational workflows

The important word here is assisted.

The appropriate level of AI autonomy, validation and human oversight should depend on the use case and its associated risk.

AI → FHIR: Can AI Help Interoperability?

The relationship can also work in the opposite direction.

Healthcare interoperability implementations involve substantial mapping, terminology, documentation and validation work.

AI could potentially assist with some of that complexity.

AI-Assisted Interoperability Implementation

Potential applications could include:

• Suggesting source-to-FHIR mappings
• Interpreting source field descriptions
• Identifying candidate resource relationships
• Highlighting ambiguous mappings
• Suggesting terminology mappings for review
• Navigating implementation guides
• Finding relevant profile requirements and constraints

I think the distinction between assistance and authority is particularly important here.

For example, an AI system may identify a plausible terminology mapping.

That doesn't necessarily make the mapping clinically or semantically correct.

Similarly, AI might suggest how a legacy field maps to a FHIR resource, while implementation and domain expertise are still required to validate whether that interpretation is appropriate.

The opportunity may therefore be less about replacing interoperability expertise and more about accelerating analysis and focusing human attention where judgment is actually required.

AI-Assisted FHIR Validation

Validation is another area I've been exploring.

FHIR already has established deterministic approaches for validating resources, profiles, terminology and conformance.

Those mechanisms remain fundamental.

But AI could potentially complement them.

Consider the difference:

FHIR / Rules-Based Validation
          ↓
Does this conform to the defined specification?

                 +

AI-Assisted Analysis
          ↓
Does something about this data,
mapping or implementation look unusual?

                 +

Human / Domain Validation
          ↓
Does this represent the intended
business or clinical meaning?

An AI capability might help identify:

• Suspicious mappings
• Semantic inconsistencies
• Unusual data patterns
• Potential implementation anomalies

But I wouldn't see AI replacing deterministic FHIR validation.

A more interesting model may be:

Deterministic validation for conformance + AI-assisted analysis for patterns and anomalies + human/domain validation for meaning.

Each layer addresses a different type of question.

Agents, Access and Guardrails

AI agents add another dimension.

An agent could potentially interact with healthcare systems through standardized interfaces—retrieving information, using tools or participating in workflows.

That creates opportunity, but it also changes the risk profile.

An AI capability that can read information is fundamentally different from one that can take actions.

And making healthcare information interoperable doesn't mean every AI capability should automatically have access to it.

A controlled architecture might look conceptually like:

Healthcare Systems
        ↓
HL7 / FHIR
        ↓
Interoperable Health Data
        ↓
Identity + Authorization
        ↓
Policy / Consent / Approved Context
        ↓
AI / RAG / Agent
        ↓
Validation + Monitoring + Human Oversight

The data may be interoperable.

Access should still be intentional.

Different AI use cases may require very different permissions, data boundaries and levels of autonomy.

Responsible AI Becomes Part of the Architecture

This is where interoperability, AI and Responsible AI begin to converge.

As AI becomes more integrated with healthcare information, considerations such as:

• Privacy
• Security
• Data quality
• Explainability
• Access control
• Human oversight
• Accountability
• Validation

can't simply be added after the solution has been built.

They increasingly become part of the architecture itself.

One way I think about the relationship is:

FHIR helps define how healthcare information can be represented and exchanged.

AI creates new possibilities for how that information can be discovered, interpreted and used.

Responsible AI and governance help establish the boundaries within which that should happen.

These are different layers of the problem, but increasingly they need to work together.

The Bigger Picture

Putting the pieces together:

Healthcare Systems
        ↓
HL7 / FHIR
        ↓
Interoperable + Meaningful Data
        ↓
Security + Governance
        ↓
AI / RAG / Agents
        ↓
Validation + Human Oversight
        ↓
Useful and Trusted Outcomes

No single layer solves the entire problem.

AI doesn't replace interoperability.

FHIR doesn't automatically make AI trustworthy.

Governance doesn't replace good architecture.

And validation doesn't end with checking whether a resource is technically conformant.

The value comes from how these layers work together.

Final Thoughts

One idea has become increasingly clear to me while exploring this intersection:

Better AI in healthcare may depend as much on better interoperability and better data foundations as it does on better models.

Models naturally receive much of the attention.

But underneath them are standards, terminology, data quality, APIs, security, validation and governance that determine what information an AI system can actually understand and use.

That's what makes the intersection of HL7, FHIR and AI particularly interesting.

Not because AI replaces healthcare interoperability.

But because interoperability may become one of the foundations that allows AI to be useful in healthcare in a more structured, governed and trustworthy way.

Perhaps the more interesting question isn't simply:

"How can AI use FHIR?"

It's:

"How should we design interoperable healthcare information so that both systems and AI can use it responsibly?"

That's an area I think is worth continuing to explore.