Describing Data Processing in FHIR: AI-Assisted Interoperability for Cancer Stage Extraction

Area of Science:

• Clinical informatics
• Artificial intelligence in healthcare
• Health data standards

Background:

• Creating interoperable clinical data models using Fast Healthcare Interoperability Resources (FHIR) is crucial but time-consuming.
• This study addresses the challenge of documenting and structuring FHIR-based data transformation workflows.

Purpose of the Study:

• To explore the application of Generative AI in automating the documentation and structuring of FHIR data transformation workflows.
• Specifically focusing on the extraction of TNM cancer staging information.

Main Methods:

• Utilized FHIR Release 4 and the PlanDefinition resource to model transformation processes.
• Employed Business Process Model and Notation (BPMN) to define workflows.
• Leveraged the large language model Claude Code (Sonnet 4.5) to generate FHIR artifacts from BPMN inputs.
• Validated generated artifacts through syntax checks, Implementation Guide compilation, and expert review.

Main Results:

• AI-assisted generation successfully produced a validated PlanDefinition with seven structured activities.
• The generated artifacts accurately represented the TNM extraction workflow.
• All FHIR artifacts were interoperable and passed conformance tests after minor revisions.

Conclusions:

• Generative AI effectively supports FHIR workflow modeling, enhancing efficiency and reducing complexity.
• Expert validation remains critical for ensuring semantic accuracy and reproducibility of AI-generated FHIR artifacts.