The AI Paradox: Why LLMs in Healthcare Actually Require More Structured Data, Not Less | Ewout Kramer & Ward Weistra (Firely)

Firely emerged from frustration with early HL7 standards, turning a decade‑long standards effort into a commercial FHIR server and the Simplifier platform. Today the company runs Dev Days, a four‑day conference that draws roughly 400 participants, over 100 sessions, and five parallel tracks. Hosted alternately in Amsterdam and Minneapolis, the event blends tutorials for newcomers with deep‑dive talks from developers, product builders, and policy experts, reinforcing the community‑first ethos that has driven FHIR’s rapid evolution.

FHIR’s rise is no longer academic; it is now embedded in national policy. In the United States, the Cures Act, along with the 21st Century Cures and upcoming HL7 FHIR‑based rules, require interoperable, patient‑accessible data. Europe follows suit, with the Dutch Ministry of Health, the European Health Data Space, and the forthcoming EGTS all mandating FHIR as the default standard. This regulatory push has turned FHIR into the de‑facto lingua franca for health‑IT integration, accelerating vendor adoption and cross‑border data exchange.

The excitement around AI in healthcare hinges on that structured foundation. Large language models can generate natural‑language queries, but without well‑defined FHIR resources they struggle to retrieve accurate clinical information. Speakers at Dev Days highlighted benchmarks that translate human questions into FHIR‑based results, emphasizing the need for robust data mapping and validation pipelines. Mis‑mapped data leads to “garbage” outputs, undermining trust in AI‑driven decision support. As more countries adopt FHIR and AI tools mature, the industry must invest in precise modeling, terminology services, and open‑source libraries to ensure that intelligent applications operate on clean, interoperable health data.