FHIR extraction pipeline.

Method at a glance

Goal: convert unstructured order text into canonical FHIR resources with low variance and high scoring reliability.

Approach

• Extract — the extractor emits compact JSON conforming to a typed contract (enums, required fields, normalized formats).
• Map — a deterministic mapper converts that JSON into FHIR resources.

Why it works

• The schema reduces ambiguity and forces completeness.
• The strict prompt contract stabilizes structure and formatting.
• The mapper enforces canonicalization and consistent displays.

Evaluation

Rubric and golden tests

• Field-level checks for presence, format, and normalization.
• Strict JSON validation against the contract.
• Deterministic comparisons on the mapped FHIR output.

Results (sanitized)

• Typical end-to-end runtimes fall under 20 seconds on internal runs.
• Perfect rubric pass across the published test set.

We share ranges only, to avoid leaking implementation details.

Operational notes

• Reproducibility: each run is tracked in CI with a run ID, commit hash, artifact names, and checksums. Public artifacts exclude environment specs and configs.
• Versioning: a change log is maintained for public materials (e.g. "v2 prompt contract tightened," "mapper enum extended").
• Redaction rules: no publication of prompts, hyper-parameters, model identifiers, device SKUs, dataset specifics, or internal paths.

Limitations & next steps

• Public artifacts are intentionally minimal.
• Certain edge cases are tracked internally and reviewed under NDA.
• The contract can adapt to additional order types and downstream FHIR resources as needs evolve.

Downloads (sanitized)

A methods-note PDF and pipeline-diagram SVG are available in sanitized form. We share prompts, configs, and full evaluation data with partners under NDA.

Interested in a deeper review? We walk partners through the full pipeline and evaluation under NDA — schedule a call.