exporting-to-fhir
GitHub将OpenMed NER提取的临床实体转换为FHIR R4资源(Condition、MedicationStatement、Observation)。利用内置导出器构建CodeableConcept并生成标准JSON,支持EHR互操作。适用于提及FHIR、特定编码或需标准化输出的场景。
Trigger Scenarios
Install
npx skills add maziyarpanahi/openmed --skill exporting-to-fhir -g -y
SKILL.md
Frontmatter
{
"name": "exporting-to-fhir",
"license": "Apache-2.0",
"metadata": {
"pairs": "after",
"project": "OpenMed",
"version": "1.0",
"category": "fhir-interop"
},
"description": "Convert OpenMed NER output (entities from openmed.analyze_text) into FHIR R4 resources — Condition, MedicationStatement, Observation — using OpenMed's built-in FHIR R4 export helpers in openmed.clinical.exporters. Covers the verified CodeableConcept builder (coding, codeable_concept, system_uri), deterministic fullUrl references, and OperationOutcome reporting. Use after running OpenMed NER when the user wants standards-conformant FHIR JSON, mentions FHIR, Condition\/Observation\/MedicationStatement, CodeableConcept, RxNorm\/LOINC\/ICD-10\/SNOMED coding, or interoperability with an EHR. Pairs after extracting-clinical-entities; feeds assembling-fhir-bundles and validating-us-core."
}
Exporting to FHIR
OpenMed's NER (openmed.analyze_text) returns spans — text, label, offsets,
confidence. To make those spans interoperable you wrap each clinically relevant
span in a FHIR R4 resource (Condition, MedicationStatement,
Observation, ...) carrying a coded CodeableConcept. OpenMed ships the
mechanical R4 export helpers for this in openmed.clinical.exporters; you
own the small amount of clinical mapping (which span becomes which resource).
When to use
Use this after NER, when the consumer is a FHIR system (an EHR, a registry, a
data lake on FHIR). Reach for it when the user says "export to FHIR", "make a
Condition/Observation", "build a CodeableConcept", or needs RxNorm/LOINC/ICD-10/
SNOMED-coded resources. For packaging many resources into one transaction
Bundle, hand off to assembling-fhir-bundles. To check the result against US
Core, hand off to validating-us-core.
What OpenMed gives you (verified API)
OpenMed deliberately ships the purely mechanical pieces and leaves clinical judgement to you. The verified entry points:
# CodeableConcept builder — openmed/clinical/exporters/codeable_concept_simple.py
from openmed.clinical.exporters.codeable_concept_simple import (
system_uri, # vocab id -> canonical HL7 system URI
coding, # (system, code, display) -> Coding dict
codeable_concept, # [Coding, ...] -> CodeableConcept dict (deterministic order)
)
# Bundle + reference + OperationOutcome — openmed/clinical/exporters/fhir/
from openmed.clinical.exporters.fhir import (
to_bundle, # [resource, ...] -> R4 transaction Bundle
deterministic_fullurl, # (doc_id, index) -> stable urn:uuid
OperationOutcomeIssue, # issue dataclass
to_operation_outcome, # [issue, ...] -> OperationOutcome
from_validation_result, # validator result -> OperationOutcome
)
system_uri knows these vocabularies out of the box: rxnorm, icd-10-cm,
loinc, snomed, hpo, mesh (and passes through any http(s):// URI
unchanged). It is the single source of truth for vocab-id → system-URI mapping.
There is no to_condition() / to_observation() magic function. You build
the resource shell (a small dict) and drop a codeable_concept(...) into its
coded slot. That is by design: the resource type and clinical status are
decisions OpenMed will not make for you.
Quick start: entity → Condition
import openmed
from openmed.clinical.exporters.codeable_concept_simple import coding, codeable_concept
# 1) NER (synthetic note — no real PHI)
result = openmed.analyze_text(
"Assessment: type 2 diabetes mellitus, stable.",
model_name="disease_detection_superclinical",
)
# result.entities -> EntityPrediction(text, label, confidence, start, end)
span = result.entities[0] # e.g. text="type 2 diabetes mellitus"
# 2) Ground to a code OUT OF PROCESS (your terminology server / mapping table).
# OpenMed bundles no restricted vocab — see querying-terminology-service.
icd_code, snomed_code = "E11.9", "44054006"
# 3) Build the CodeableConcept with OpenMed's builder
cc = codeable_concept(
[
coding("snomed", snomed_code, "Diabetes mellitus type 2"),
coding("icd-10-cm", icd_code, "Type 2 diabetes mellitus without complications"),
],
text=span.text,
)
# 4) Assemble the resource shell yourself
condition = {
"resourceType": "Condition",
"id": "cond-1",
"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",
}]
},
"category": [{
"coding": [{
"system": "http://terminology.hl7.org/CodeSystem/condition-category",
"code": "encounter-diagnosis",
}]
}],
"code": cc, # OpenMed-built CodeableConcept
"subject": {"reference": "Patient/patient-1"},
"recordedDate": "2024-03-02",
}
codeable_concept sorts codings deterministically (SNOMED, LOINC, RxNorm,
ICD-10-CM, HPO, MeSH first; everything else alphabetical), so the JSON is
byte-stable across runs — important for diffable pipelines and golden tests.
Worked: the resulting Condition.code
{
"code": {
"coding": [
{ "system": "http://snomed.info/sct", "code": "44054006",
"display": "Diabetes mellitus type 2" },
{ "system": "http://hl7.org/fhir/sid/icd-10-cm", "code": "E11.9",
"display": "Type 2 diabetes mellitus without complications" }
],
"text": "type 2 diabetes mellitus"
}
}
Workflow
- NER —
openmed.analyze_text(note, model_name=...)→result.entities. - Classify each span: a disease label →
Condition; a drug →MedicationStatement; a lab/vital/measurement →Observation. - Ground the surface text to a code out of process (terminology server or
your own map). Never invent codes; if you cannot ground a span, emit a
CodeableConceptwith onlytextand nocoding. - Build the
CodeableConceptwithcoding(...)+codeable_concept(...). - Wrap it in the resource shell (set
clinicalStatus/status,subject, dates). Use the cheat-sheet below. - Reference the Patient/Encounter via
{"reference": "Patient/<id>"}. - Pass the list to
to_bundle(...)(assembling-fhir-bundles) and validate (validating-us-core).
Resource cheat-sheet (where the CodeableConcept goes)
| OpenMed entity kind | FHIR resource | Coded slot | Required status field |
|---|---|---|---|
| Disease / diagnosis | Condition |
code |
clinicalStatus, verificationStatus |
| Drug / medication | MedicationStatement |
medicationCodeableConcept |
status (e.g. active) |
| Lab / vital / finding | Observation |
code (+ valueQuantity/valueCodeableConcept) |
status (e.g. final) |
| Procedure | Procedure |
code |
status |
| Allergy | AllergyIntolerance |
code |
clinicalStatus |
MedicationStatement (drug span)
med = {
"resourceType": "MedicationStatement",
"id": "med-1",
"status": "active",
"medicationCodeableConcept": codeable_concept(
[coding("rxnorm", "860975", "metformin hydrochloride 500 MG Oral Tablet")],
text="metformin 500 mg",
),
"subject": {"reference": "Patient/patient-1"},
}
Observation (lab/vital span)
obs = {
"resourceType": "Observation",
"id": "obs-1",
"status": "final",
"category": [{"coding": [{
"system": "http://terminology.hl7.org/CodeSystem/observation-category",
"code": "laboratory",
}]}],
"code": codeable_concept(
[coding("loinc", "4548-4", "Hemoglobin A1c/Hemoglobin.total in Blood")],
text="HbA1c",
),
"valueQuantity": {
"value": 7.4, "unit": "%",
"system": "http://unitsofmeasure.org", "code": "%",
},
"subject": {"reference": "Patient/patient-1"},
}
Hand-off to / from OpenMed
- From OpenMed:
result.entities(EntityPrediction.text/.label/.confidence /.start/.end) is the input. Keepconfidenceand the offsets in an extension or a side log so the resource is auditable back to the source span. - To OpenMed: before exporting a note that still contains PHI, run
openmed.deidentify(...); or de-identify a built resource/Bundle withopenmed.interop.fhir_operations.de_identify_resource/de_identify_bundle(see that module — it walks free-text + narrative and never touches codes, references, systems, or temporal values). - OperationOutcome: report any spans you could not map as
OperationOutcomeIssue(severity="warning", code="incomplete", diagnostics=..., expression="Condition.code")→to_operation_outcome([...]). Keep diagnostics PHI-free (offsets/labels, never raw identifiers).
Edge cases & gotchas
- No code? Still valid. A
CodeableConceptwith onlytextand nocodingis legal R4. Emit it rather than inventing a code, and flag it via OperationOutcome. US Core may still require a code — seevalidating-us-core. system_uriraises on an unknown short id that is not a URL. Pass a known short id (rxnorm/loinc/snomed/icd-10-cm/hpo/mesh) or a fullhttp(s)://system URI.- Negation / temporality. OpenMed NER finds the mention, not its assertion.
A negated ("no diabetes") or historical span should change
verificationStatus/clinicalStatusor be dropped. Resolve assertion first (resolving-clinical-context,openmed.clinical). - Stable ids. Give each resource a unique
id;to_bundlerejects duplicateResourceType/idpairs because they corrupt cross-references. - Local-first. Grounding to RxNorm/LOINC/SNOMED is out-of-process with the user's own credentials. OpenMed bundles no restricted terminology.
Standards & references
- FHIR R4 Condition: https://hl7.org/fhir/R4/condition.html
- FHIR R4 MedicationStatement: https://hl7.org/fhir/R4/medicationstatement.html
- FHIR R4 Observation: https://hl7.org/fhir/R4/observation.html
- FHIR R4 CodeableConcept: https://hl7.org/fhir/R4/datatypes.html#CodeableConcept
- HL7 terminology systems (system URIs): https://hl7.org/fhir/R4/terminologies-systems.html
- FHIR R4 OperationOutcome: https://hl7.org/fhir/R4/operationoutcome.html
Version History
- f213557 Current 2026-07-23 00:44


