Context-Free Grammar-Guided Generation of FHIR Resources Using Large Language Models

Area of Science:

• Computational linguistics
• Health informatics
• Artificial intelligence

Background:

• Large Language Models (LLMs) excel at medical data tasks but produce unstructured outputs.
• Integrating LLM outputs into clinical data pipelines is challenging due to lack of structure.
• Fast Healthcare Interoperability Resources (FHIR) offers standardized medical data exchange but is complex to generate.

Purpose of the Study:

• To develop a method for generating schema-conformant structured data using LLMs.
• To address the limitations of LLM unstructured outputs in clinical data integration.
• To enable reliable generation of complex FHIR resources.

Main Methods:

• Developed a pipeline to synthesize Context-Free Grammars (CFGs) from FHIR JSON Schemas.
• Employed relaxed key ordering in CFG generation for structural validity.
• Evaluated CFG-guided LLM generation against JSON Schema-guided and unguided methods using FHIR R5 resources.

Main Results:

• CFG-guided LLM generation achieved 76.5% validity for FHIR resources, significantly outperforming JSON Schema-guided (41.2%) and unguided (17.6%) methods.
• The approach successfully enforced value constraints, including enumerated types and UCUM unit codes.
• Demonstrated superior structural compliance compared to other generation methods.

Conclusions:

• Synthesizing CFGs from FHIR schemas is an effective strategy for schema-conformant structured data generation with LLMs.
• This method enhances the reliability and validity of LLM-generated healthcare data for clinical applications.
• The open-source release of the grammar synthesis pipeline aims to promote adoption in healthcare information systems.