Parsing C-CDA / CCD Documents for OpenMed

SkillDocs & knowledge

Lets your agent read C-CDA clinical documents and extract readable sections plus coded medical entries.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Parsing C-CDA / CCD Documents for OpenMed skill

About this capability

Parses C-CDA / CCD XML clinical documents to extract human-readable section narrative plus coded entries, keyed by section LOINC codes and templateIds. Use before OpenMed processing when ingesting C-CDA R2.1 documents (CCD, Discharge Summary, H&P, Consultation Note) exported from an EHR and you need

What this skill tells your AI

The instructions your AI receives, as published by maziyarpanahi/openmed in skills/parsing-ccda-documents/SKILL.md and read by ahel’s review.

C-CDA (Consolidated Clinical Document Architecture) is the XML document standard behind Meaningful Use / ONC certification — the CCD, Discharge Summary, History & Physical, and Consultation Note you get when an EHR "exports a chart". Each document is a ClinicalDocument with a header (patient, authors, encounter) and a structuredBody of sections. Every section has two representations: a human-readable narrative <text> block and machine-readable coded entries. The narrative is what you feed to clinical NLP. This skill extracts it and hands it to OpenMed.

When to use

  • You receive C-CDA R2.1 / CCD documents (Direct messaging, patient portal export, HIE) and want the free-text section narrative for de-id and NER.
  • You need to pair narrative spans with the section they came from (problems, meds, allergies, results, plan, H&P narrative).
  • You want XML-safe de-identification that keeps the document parseable.

C-CDA structure in one minute

<ClinicalDocument xmlns="urn:hl7-org:v3">
  <recordTarget><patientRole>
    <id extension="12345" root="..."/>
    <patient><name><given>Jane</given><family>Doe</family></name>
      <birthTime value="19700115"/></patient>
  </patientRole></recordTarget>
  <component><structuredBody>
    <component><section>
      <templateId root="2.16.840.1.113883.10.20.22.2.5.1"/>   <!-- Problems -->
      <code code="11450-4" codeSystem="2.16.840.1.113883.6.1"/> <!-- LOINC -->
      <title>Problems</title>
      <text>Active problems: Type 2 diabetes, hypertension.</text>  <!-- narrative -->
      <entry>...coded SNOMED/ICD entries...</entry>
    </section></component>
  </structuredBody></component>
</ClinicalDocument>

Sections are identified by templateId/@root and by section code (LOINC). The CDA namespace is urn:hl7-org:v3.

Quick start

Extract section narrative by LOINC code, then hand off to OpenMed:

import openmed
from xml.etree import ElementTree as ET

NS = {"hl7": "urn:hl7-org:v3"}
SECTION_LOINC = {
    "11450-4": "problems", "10160-0": "medications", "48765-2": "allergies",
    "30954-2": "results",  "18776-5": "plan",        "10164-2": "hpi",
    "8648-8": "hospital_course", "11488-4": "consult_note",
}

root = ET.parse("ccd.xml").getroot()
for section in root.findall(".//hl7:section", NS):
    code_el = section.find("hl7:code", NS)
    loinc = code_el.get("code") if code_el is not None else None
    text_el = section.find("hl7:text", NS)
    if text_el is None:
        continue
    narrative = "".join(text_el.itertext()).strip()       # flatten narrative block
    if not narrative:
        continue

    deid = openmed.deidentify(narrative, method="replace", policy="hipaa_safe_harbor")
    result = openmed.analyze_text(deid.text, output_format="dict")
    section_name = SECTION_LOINC.get(loinc, loinc)
    # attach (section_name, result) for downstream consumers

"".join(text_el.itertext()) flattens the narrative block (which may contain <paragraph>, <list>, <table>, <content> markup) into plain text.

XML-aware whole-document de-identification

When you need to redact PHI from the document (header ids, names, addresses, dates) while keeping the CDA XML valid and parseable, use the bundled adapter rather than regexing the raw XML:

from openmed.interop.cda import redact_cda, is_cda_document

if is_cda_document("ccd.xml"):
    safe_xml = redact_cda("ccd.xml")     # returns redacted XML string

redact_cda applies DEFAULT_PHI_ELEMENT_MAP (patient id hashed, name/address/ telecom null-flavored, birthTime and effectiveTime date-shifted) to header elements and sweeps section narrative text — operating on text nodes only so surrounding markup stays intact. Pass text_redactor= to plug an extra free-text callback (e.g. an openmed.deidentify wrapper), date_shift_days= for a fixed shift, and keep_year=True to preserve years.

Workflow

  1. Confirm it's CDA. is_cda_document(...) checks for a ClinicalDocument root. Reject XML with DOCTYPE/ENTITY declarations (XXE risk) — the adapter does this for you.
  2. Read the header for context: patient, author, effectiveTime, documentType (ClinicalDocument/code LOINC). Treat all header values as PHI.
  3. Walk sections by templateId or section code (LOINC). Map to your section vocabulary.
  4. Flatten narrative <text> with itertext(); preserve the section→text association for span attribution.
  5. De-identify → analyze each narrative with OpenMed. Prefer coded <entry> data when it already exists; use NLP to recover what is only in narrative.

Hand-off to / from OpenMed

  • To OpenMed: flattened section narrative → openmed.deidentifyopenmed.analyze_text. Keep (section LOINC, narrative) so entities trace back to their section.
  • Adapter: openmed.interop.cda provides redact_cda, is_cda_document, PhiElementRule, and DEFAULT_PHI_ELEMENT_MAP for namespace-aware, markup-preserving de-identification. It also registers an .xml document handler with OpenMed's multimodal intake, so .xml files are auto-detected as CDA and redacted on ingest.
  • Onward: re-emit findings via openmed.clinical.exporters.fhir or align narrative-derived problems to the section's coded entries.

Edge cases & gotchas

  • Narrative vs entries can disagree. The human-readable <text> is authoritative for display, coded <entry> for machines — they sometimes drift. Reconcile, and prefer narrative for what NLP must recover.
  • <content ID=...>/<reference> linkage. Narrative <content> elements carry IDs referenced by entries (<reference value="#problem1"/>); use them to link a coded entry to its exact narrative phrase.
  • Tables and lists. Section narrative often uses <table>/<list>; itertext() flattens these — re-impose structure if column meaning matters.
  • Namespaces & prefixes. Always bind the urn:hl7-org:v3 namespace; some documents add sdtc: extensions and xsi: typing.
  • XXE / unsafe XML. Never parse untrusted CDA with entity expansion enabled; the adapter rejects DOCTYPE/ENTITY outright — do the same in custom parsers.
  • Restricted terminology. Coded entries reference SNOMED CT, RxNorm, LOINC; OpenMed does not bundle SNOMED/CPT — resolve codes against the user's own licensed terminology out-of-process.

Standards & references

Signals

GitHub stars
5k
Forks
666
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
parsing-ccda-documents
Source
github.com/maziyarpanahi/openmed