extract-clinical-entities-to-fhir

Extract clinical entities from synthetic or already de-identified text with OpenMed and map them into deterministic FHIR R4 resources and a Bundle. Use when an agent must turn local clinical NER outpu

SKILL.md

GitHub

Extract clinical entities to FHIR

Separate extraction from clinical coding. OpenMed finds spans and supplies the mechanical FHIR builders; the application decides which resource type and status are clinically appropriate.

Procedure

  1. Keep the source synthetic, or de-identify it inside the trusted boundary before extraction.
  2. Run openmed.analyze_text with the task-appropriate clinical model.
  3. Filter predictions by label and confidence; preserve offsets in a PHI-safe audit record.
  4. Map each accepted span to the correct FHIR resource type.
  5. Add terminology codes only from a user-approved mapping or terminology service. Never invent a code.
  6. Assemble resources with to_bundle and validate against the target profile.

Runnable synthetic example

Install the model runtime first with python -m pip install "openmed[hf]".

import json

from openmed import analyze_text
from openmed.clinical.exporters.fhir import to_bundle

note = "Assessment: type 2 diabetes mellitus is stable on metformin."
result = analyze_text(
    note,
    model_name="disease_detection_superclinical",
    confidence_threshold=0.5,
)

resources = [{"resourceType": "Patient", "id": "synthetic-patient"}]
for index, entity in enumerate(result.entities, start=1):
    if entity.label.upper() not in {"CONDITION", "DIAGNOSIS", "DISEASE"}:
        continue
    resources.append(
        {
            "resourceType": "Condition",
            "id": f"condition-{index}",
            "clinicalStatus": {
                "coding": [
                    {
                        "system": (
                            "http://terminology.hl7.org/CodeSystem/"
                            "condition-clinical"
                        ),
                        "code": "active",
                    }
                ]
            },
            "verificationStatus": {
                "coding": [
                    {
                        "system": (
                            "http://terminology.hl7.org/CodeSystem/"
                            "condition-ver-status"
                        ),
                        "code": "confirmed",
                    }
                ]
            },
            # A text-only CodeableConcept is preferable to an invented code.
            "code": {"text": entity.text},
            "subject": {"reference": "Patient/synthetic-patient"},
        }
    )

if len(resources) == 1:
    raise RuntimeError("No condition spans met the label and confidence rules")

bundle = to_bundle(resources, doc_id="synthetic-note-001")
print(json.dumps(bundle, indent=2))

Safety checks

  • Do not put raw identifiers, source text, or reversible mappings in logs, OperationOutcome.diagnostics, or trace metadata.
  • Keep a patient identity service separate from extracted clinical facts.
  • Preserve negation, temporality, and experiencer context before asserting a resource as active or confirmed.
  • Use a text-only CodeableConcept when no approved code is available.
  • Validate the Bundle against the receiver's FHIR and profile requirements.
  • Do not bundle restricted terminologies; use the user's licensed service.

Repository example

Read and run the redaction-to-FHIR walkthrough for an offline-friendly pipeline with deterministic extraction.