Describing Data Processing in FHIR: AI-Assisted Interoperability

Authors: David OUAGNE¹, Vincent ZOSSOU¹, and Bastien RANCE^1,2

Affiliations:

• ¹ AP-HP, Paris, France
• ² Centre de Recherche des Cordeliers, UMRS 1138, Université Paris Cité, Inserm, Sorbonne Université, Paris, France

ORCID IDs:

• David Ouagne: 0009-0001-4069-6124
• Vincent Zossou: 0000-0001-7016-4455
• Bastien Rance: 0000-0003-4417-1197

---

Summary

This study demonstrates the use of AI-assisted FHIR Implementation Guide authoring for documenting complex data transformation workflows in healthcare. Using TNM cancer staging as a use case, we explored how FHIR PlanDefinition resources can formalize the computational transformation process from raw clinical data to structured, actionable FHIR resources.

Key Contributions

• Novel Use of PlanDefinition: We repurposed the FHIR PlanDefinition resource—traditionally used for care plans—to document and structure data transformation pipelines, representing sequences, dependencies, and conditions in analytical workflows.

• Five-Category Framework: Introduced a concise typology for organizing transformation processes: Structuration, Extraction, Normalization, Consolidation, and Derivation.

• BPMN-to-FHIR Translation: Developed a workflow using Business Process Model and Notation (BPMN) as a human-readable blueprint, which was then translated into formal FHIR PlanDefinition artifacts.

• AI-Assisted Generation: Leveraged Claude Code (Sonnet 4.5) to translate BPMN models into FHIR-compliant PlanDefinition resources in FSH format, requiring only 1-2 rounds of refinement to pass all validation controls.

Methods

The approach involved creating BPMN diagrams to map the TNM staging workflow, documenting data sources and transformation steps, and using a structured prompting protocol to guide Claude Code in generating FHIR PlanDefinition resources. The generated artifacts were validated through syntax checking, Implementation Guide compilation, and expert review.

Results & Impact

This work illustrates that AI can effectively translate conceptual models into formal FHIR structures, simplifying artifact generation, reducing costs in time and expertise, and assisting non-experts in adopting FHIR. The approach demonstrates how Generative AI can serve as a collaborative assistant in achieving transparent, reproducible, and semantically rich data transformation pipelines.

BPMN used to represent the data extraction and integration workflow

TNM BPMN file

---

FHIR Implementation Guide

The complete FHIR Implementation Guide with the generated PlanDefinition for TNM Cancer Staging is available at:

CARPEM Oncology Staging PlanDefinition

---

Claude Code Prompt

The complete Claude Code Prompt used to generate the PlanDefinition artifact is available at:

Claude Code Prompt

The resulting PlanDefinition

The resulting PlanDefinition artifact (as a FHIR FSH file) is available at:

TNM PlanDefinition Fsh file

Reproductibility study

We produce five versions of the PlanDefinition using the identical context and prompt. The files are available in the repro_study folder. The full analysis of the differences among the five PlanDefinition versions is available at: Reproducibility study

Contact

For more information or to cite this work, please contact: David Ouagne: david.ouagne@aphp.fr