After decades of working toward seamless health data interoperability, we find ourselves at another pivotal moment. The rapid adoption of Artificial Intelligence (AI) in healthcare presents us with a familiar challenge wearing a new face: how do we ensure these powerful new tools work together transparently, accountably, and in the service of better health for people everywhere?
At a recent ADAPT conference panel, I had the opportunity to reflect on what our interoperability journey can teach us as we venture into standardizing intelligence, not just data. Here are some key insights from that conversation.
First, a grounding perspective: this is a journey, not a destination. Despite all the progress we've made in healthcare interoperability, too often, people still move faster and further than their health information. The ability for any digital tool—including AI—to help people make better health decisions is always limited by the scope of data in its purview and its capability to make sense of it.
Even the most powerful AI we can imagine must overcome the same boundaries we've always faced: technical, organizational, business, and jurisdictional barriers that prevent us from seeing the complete picture of health information relevant for individuals or populations.
However, HL7's decade-plus journey with Fast Health Interoperability Resources (FHIR® ) has taught us something crucial: open standards are a potent fuel for innovation. The vibrant, open, collaborative community around FHIR wasn't just a nice byproduct—it was the key force that created a well-tuned specification and enabled it to flourish in the marketplace.
Open standards level the playing field, reduce barriers to participation, and free organizations from proprietary formats. They unlock new connectivity, preserve data sovereignty, and most fundamentally, enable new digital freedoms. As we approach AI standardization, maintaining this commitment to openness isn't guaranteed, but it's the future we're fighting for.
Forward thinkers posed the question: if FHIR gave us a common data language, what's the equivalent for AI models?
Perhaps we should call it FHLAIM🔥—Fast Healthcare Language for AI Models! Mostly kidding.
As we consider the types of interoperability approaches required for the AI era, our foundation is solid. FHIR provides an excellent framework for building on and extending. Combined with HL7's adjacent standards, such as FHIR Bulk Data Access, SMART App Launch, CDS Hooks, and Clinical Quality Language (CQL), we have a versatile toolkit for driving innovation with AI. In HL7's recent AI Challenge, I was blown away by the range of standards-powered AI innovations from startups to established companies alike that leveraged these capabilities in novel ways.
In the new AI Transparency on FHIR project, the HL7 community is developing technical guidance for when AI uses and influences health data with FHIR. Some aspects of this work include:
As we move into the era of conversational AI and agent-based systems, we'll need fine-tuned specifications for representing what AI model tools are, what they can do, and how to define their expected inputs and outputs. I expect this to be an active area of exploration for us, including how industry specifications like Model Context Protocol (MCP) and Agent2Agent (A2A) Protocol can be adapted for use in conjunction with FHIR-based approaches.
Health systems are rapidly implementing AI. Some academic medical centers now deploy 200+ AI models, yet many organizations lack standardized approaches to validation, bias detection, fairness evaluation across diverse populations, and real-time performance surveillance. The scale of deployment, diversity of solutions, and the dynamic post-deployment performance of AI systems necessitate new institutional governance mechanisms. HL7's health system partners are initiating a conversation about the open standards that could enable real-time, shareable, and distributed monitoring of AI performance and fairness across the healthcare ecosystem. A distributed monitoring framework will be key to scaling our AI governance approaches.
Standards developing organizations like HL7 play a critical role in enabling trustworthy AI. First, we serve as a trusted, neutral convener, bringing together the diverse range of stakeholders necessary for these conversations, including delivery organizations, patients, technology vendors, payers, regulators, and all others with an interest in getting this right.
Other organizations convene stakeholders, too. What makes HL7 unique is our proven governance model that consistently turns consensus into actionable public goods, including open, freely available technical specifications for bringing those concepts to life in IT systems. Our role is to transform industry consensus that says "AI decisions need to be traceable, auditable, and transparent" into "Here's the technical standard that makes those features portable and consistent across systems." HL7's standards development process leverages global interoperability expertise to create precise technical guidance on how AI systems should function and integrate.
At the same time, it's equally important to recognize what we don't do: we're not de facto regulators. We describe how to enable features like traceability and auditability, but it's up to others to decide how that information should be used.
Can you imagine a world where AI model provenance and explainability are standardized and ubiquitous? I do. In fact, I think it's inevitable.
The real question is whether we build this infrastructure proactively or reactively after a catastrophic failure.
Historically, there hasn't been a great deal of enthusiasm for standardizing data provenance. Pretty much everyone agrees we need good data provenance; it's just that they implement their own approaches internally. For the most part, there haven't been strong use cases or business drivers for standardization across the ecosystem. AI is changing that calculation, where standardized data provenance is becoming increasingly appreciated. I see the consensus expectations for model provenance and explainability converging even faster. As the scale and diversity of model solutions explode, demand for standardized approaches will follow.
I'll be honest: it will likely be an uneven and somewhat messy process. But here's why I'm confident it will happen: the alternative is unacceptable. Deploying AI systems in healthcare at scale without standardized explainability and provenance is akin to building airplanes without flight recorders. Eventually, something will go badly wrong, and we will wish we'd built the infrastructure sooner.
As we plan to scale our use of AI, we must also consider the infrastructure that will enable us to govern and monitor AI with agility and accountability. The question isn't whether we get there, it's whether we arrive before or after the crisis that forces our hand.
So what does a trustworthy, transparent, and accountable AI ecosystem in healthcare look like by 2035? Predictions for 10 years ahead in "AI time" are quite speculative. But the kind of future I want to create is one where ubiquitous, standards-powered AI comes with the features that engender trust. These are the kinds of systems that will enable new digital freedoms and open opportunities for everyone to live in optimal health. Getting there requires the same collaborative spirit, commitment to openness, and patient persistence that brought us FHIR. The challenges are different, but the principles remain the same: standards start with and fundamentally serve people.
The journey continues.