Taiwan's health ministry says three hospital groups completed a cross-system Fast Healthcare Interoperability Resources (FHIR) demonstration by the end of 2025. The project adds a common data layer above existing hospital information systems rather than requiring every institution to replace its core software.
FHIR can give applications and artificial intelligence systems a more predictable way to find patient, laboratory and medication data. It cannot guarantee that the source records are complete or correctly coded, and it does not decide which users or automated agents may see them.
FHIR turns records into reusable resources
Health Level Seven International (HL7), the organization that develops the standard, expands FHIR as Fast Healthcare Interoperability Resources. HL7 describes FHIR systems as collections of modular Resources that can represent administrative and clinical concepts, including patients, organizations, medications and diagnostics.
A Resource has defined fields and can be adapted through profiles and extensions for a particular use. The standard uses familiar web technologies, including XML, JSON and HTTP, and supports RESTful exchanges. That structure lets an application request a specific type of information instead of treating an entire clinical document as an undifferentiated block of text.
Structured data changes the AI task
General-purpose language models can return plausible but incorrect medical codes. A 2024 NEJM AI study tested GPT-4, GPT-3.5, Gemini Pro and Llama 2-70B against more than 27,000 diagnosis and procedure codes; GPT-4 produced the highest exact-match rates, including 33.9 percent for ICD-10-CM and 49.8 percent for CPT. The researchers prompted the models with code descriptions, so the experiment was a code-query benchmark rather than a test of models reading complete patient charts.
A separate 2024 preprint examined a different workflow. Its authors used FHIR Resources and an oncology profile called Minimal Common Oncology Data Elements (mCODE), reporting 87 percent accuracy for SNOMED CT, 90 percent for LOINC and 84 percent for RxNorm. They also reported overall data-standardization accuracy above 92 percent.
Those percentages do not show that converting a record to FHIR automatically lifts coding accuracy from about one-third to about 90 percent. The papers used different inputs, models, tasks and evaluation methods. The mCODE paper was also a preprint, and its results do not establish clinical safety or improved patient outcomes.
Taiwan is layering FHIR over existing hospital systems
Taiwan's Ministry of Health and Welfare, which oversees the country's health system, calls its shared layer FHIR Box. The ministry says the platform combines data conversion, storage and exchange, while Taiwan Core Data for Interoperability and a national implementation guide define how local systems represent information.
The ministry reported that the Chang Gung, Mackay and Chung Shan hospital groups had completed a cross-system record-sharing demonstration by the end of 2025. The announcement presents a completed demonstration, not proof that every hospital in Taiwan can already retrieve a complete record from every other provider.
A separate government program offers a measurable view of adoption in one workflow. Taiwan's National Health Insurance Administration runs the country's national health insurance program. The agency said that 81 hospitals were participating in electronic cancer-drug prior authorization by April 2026 and that 35 of them were submitting records in FHIR format.
The agency also began a next-generation sequencing data project in October 2025 because genetic test reports had often arrived as PDFs or unstructured text with inconsistent formats. Its English-language announcement says the project uses FHIR to standardize data collection across hospitals and testing institutions. That project covers a defined insurance and cancer-data workflow; it is not a count of hospitals with comprehensive FHIR deployment.
An interface standard is not a security system
FHIR determines how systems represent and exchange information. HL7 states that FHIR is not a security protocol and assumes that a separate system handles authentication, access-control decisions and audit logs. HL7 recommends protected transport, authenticated clients and OAuth-based authorization where appropriate, while providing resources such as AuditEvent that implementations can use to record access.
This separation matters when software can make many requests without a person opening each record. Hospitals must decide whether an application can read, create, update or delete a given Resource, how consent affects a token, and whether searches expose related records. A technically valid FHIR response can still violate policy if the requesting system received excessive permissions.
AI agents add another access path
The Model Context Protocol (MCP) is an open-source standard that lets AI applications connect to external data sources, tools and workflows. A developer can place an MCP server in front of a FHIR service so an agent can request structured operations instead of receiving a whole record as free text.
That design does not remove the controls required at the FHIR service. The system still needs narrowly scoped credentials, authorization for each operation, protected logs and a way to distinguish the human user from the automated client acting for that person. APPI News could not find an independent, published evaluation showing how FHIR access through MCP performs under routine hospital workloads at the time of writing.
What remains unverified
Taiwan's official pages support the existence of FHIR Box, the three-group demonstration and the participation figures for the cancer-drug authorization program. APPI News could not verify from an official publication the source article's schedule for deployment across every medical center and smaller hospital, its FHIR Box throughput figures, or its exact percentages for mapping Taiwan's local codes to LOINC and RxNorm.
The available evidence therefore supports a narrower conclusion: Taiwan has moved FHIR from standards work into multi-hospital demonstrations and a live insurance submission process. Whether that foundation produces reliable medical AI will depend on the quality of each converted Resource, the scope of each implementation and the controls placed around every application that reads or changes the data.
Sources and further reading
- Introducing HL7 FHIR(HL7 International)
- AI Falls Short: Large Language Models Struggle With Medical Coding, Study Shows(Mount Sinai via SciTechDaily)
- Novel Development of LLM Driven mCODE Data Model for Improved Clinical Trial Matching(arXiv)
- Taiwan uses FHIR technology to build health-data interoperability(Taiwan Ministry of Health and Welfare)
- The NHIA adopts FHIR standards for NGS data collection(Taiwan National Health Insurance Administration)
- FHIR security(HL7 International)
- What is the Model Context Protocol?(Model Context Protocol)