Exporting Bulk FHIR

When you need cohort-scale clinical text — not one patient in a UI — you use the FHIR Bulk Data Access ($export) operation: an async job that emits NDJSON files of resources you then stream into OpenMed for batch de-identification and NER. This skill sits before the OpenMed pipeline: it is how the notes arrive.

When to use

Reach for it when the source is an EHR or FHIR data warehouse and the volume is a population/group (thousands of patients), the workload is headless (no clinician UI), and the goal is to batch-feed openmed.deidentify / openmed.analyze_text. Triggers: "bulk export", "$export", "NDJSON", "Flat FHIR", "cohort de-identification", "export all notes". For a single in-chart patient with a UI, use scaffolding-smart-on-fhir instead.

Three export levels

• System — GET [base]/$export — everything the client is authorized for.
• Group — GET [base]/Group/[id]/$export — a defined cohort (most common).
• Patient — GET [base]/Patient/$export — all patients in scope.

Bulk export uses SMART Backend Services auth (a system/*.read-scoped client-credentials token via a signed JWT assertion), not an interactive launch.

Quick start: kickoff → poll → download

# 1) Kickoff (async). Ask for clinical-note-bearing resource types.
curl -s -X GET \
  'https://ehr.example/fhir/Group/cohort-42/$export?_type=DocumentReference,DiagnosticReport&_since=2024-01-01T00:00:00Z' \
  -H 'Authorization: Bearer <backend-services-token>' \
  -H 'Accept: application/fhir+json' \
  -H 'Prefer: respond-async' -D -
# -> 202 Accepted
#    Content-Location: https://ehr.example/fhir/bulkstatus/JOB123

# 2) Poll the status URL until complete
curl -s 'https://ehr.example/fhir/bulkstatus/JOB123' \
  -H 'Authorization: Bearer <token>'
# 202 + X-Progress while running; 200 + a manifest JSON when done:
# { "transactionTime": "...", "request": "...", "requiresAccessToken": true,
#   "output": [
#     { "type": "DocumentReference",
#       "url": "https://ehr.example/fhir/bulkfiles/dr-1.ndjson" },
#     { "type": "DiagnosticReport",
#       "url": "https://ehr.example/fhir/bulkfiles/dx-1.ndjson" } ] }

# 3) Download each NDJSON file (one FHIR resource per line)
curl -s 'https://ehr.example/fhir/bulkfiles/dr-1.ndjson' \
  -H 'Authorization: Bearer <token>' -o dr-1.ndjson

Key headers/params: Prefer: respond-async (required to start the job), Content-Location (the status/polling URL), _type (limit resource types), _since (incremental export), _typeFilter (server-side resource filtering). Delete the job when done: DELETE <status-url>.

Stream NDJSON into OpenMed (batch)

NDJSON is one resource per line — stream it; do not load the whole file. Pull the note text out of each DocumentReference/DiagnosticReport and run OpenMed on-device, in batch:

import base64, json, openmed

def note_text(resource: dict) -> str | None:
    # DocumentReference.content[].attachment.data (base64) or .url -> Binary
    for content in resource.get("content", []):
        att = content.get("attachment", {})
        if att.get("data"):
            return base64.b64decode(att["data"]).decode("utf-8", "replace")
    # DiagnosticReport.presentedForm[].data
    for form in resource.get("presentedForm", []):
        if form.get("data"):
            return base64.b64decode(form["data"]).decode("utf-8", "replace")
    return None

with open("dr-1.ndjson", "r", encoding="utf-8") as fh:
    for line in fh:                              # streaming, line by line
        resource = json.loads(line)
        text = note_text(resource)
        if not text:
            continue
        # De-identify every note before anything downstream sees it
        deid = openmed.deidentify(text, method="replace", policy="hipaa_safe_harbor")
        # Then NER on the de-identified text
        entities = openmed.analyze_text(
            deid.text, model_name="disease_detection_superclinical")
        # ... persist de-identified text + spans; never persist raw PHI

For large cohorts, parallelise across files (each NDJSON file is independent) and reuse a single OpenMed model loader across notes to avoid reloading weights.

Workflow

1. Obtain a SMART Backend Services token (system/DocumentReference.read, etc.).
2. Kickoff $export at the right level with _type (and _since for incrementals) + Prefer: respond-async.
3. Poll Content-Location until 200; read the manifest output[].
4. Download each NDJSON file (send the token if requiresAccessToken).
5. Stream each line → extract note text → openmed.deidentify → openmed.analyze_text.
6. Export findings to FHIR if needed (exporting-to-fhir, assembling-fhir-bundles).
7. DELETE the bulk job to free server storage.

Hand-off to / from OpenMed

• Into OpenMed (the point of this skill): NDJSON note text → batch openmed.deidentify is the primary hand-off. De-identify first; treat every exported note as PHI until it has been through the de-id pass.
• Back to FHIR: the spans from analyze_text → exporting-to-fhir → to_bundle; write back only if your governance allows.
• Local-first at scale: OpenMed runs on-device, so the cohort never leaves your infrastructure for NLP. Only the export traffic touches the EHR.

Edge cases & gotchas

• It's async — never block on the kickoff. A 202 + Content-Location is success; poll with backoff and honour Retry-After/X-Progress.
• Files can be huge. Stream NDJSON line-by-line; do not json.load a whole file. Parallelise per file, not per line.
• requiresAccessToken. If the manifest says so, send the bearer token when downloading the NDJSON files too.
• De-identify before persistence. Raw exported notes are PHI; the first durable artifact must be de-identified. Verify de-id with openmed.eval leakage gates (evaluating-with-leakage-gates), not F1 alone.
• Note formats vary. Text may be inline base64, an external Binary reference, or RTF/HTML in presentedForm. Normalise to plain text before OpenMed; for scanned PDFs use OpenMed's document/OCR intake.
• Clean up the job. Servers may cap concurrent/stored exports; DELETE the status URL when finished.
• Scope minimally. Request only the resource types you will process; honour the cohort's consent/governance.

Standards & references

• FHIR Bulk Data Access (Flat FHIR) IG: https://hl7.org/fhir/uv/bulkdata/
• $export operation: https://hl7.org/fhir/uv/bulkdata/export.html
• Async request pattern: https://hl7.org/fhir/R4/async.html
• SMART Backend Services auth: https://hl7.org/fhir/uv/bulkdata/authorization/
• NDJSON: https://github.com/ndjson/ndjson-spec