Apex Atlas — Synthetic FHIR patient populations

Production mapping

Not just conformance — data that maps to what you ingest in the field

Atlas output uses the same FHIR resource graph as clinical and payer channels. Population statistics mirror public US epidemiology; individual records are fully synthetic.

Channel	Atlas mirrors
QHIN / TEFCA	US Core clinical summaries — problems, meds, encounters, vitals/labs
HIE	Longitudinal multi-condition charts, notes (DocumentReference), cross-org care patterns
Stedi / claims APIs	Coverage, Claim, ExplanationOfBenefit (`--with-coverage --with-claims`)
CMS Blue Button 2.0	Medicare-shaped Coverage + EOB, age-stratified chronic care
Bulk `$export`	NDJSON + Parquet at scale

Atlas does not ingest from these systems. It produces license-clean cohorts your pipelines can treat like production FHIR — without PHI.

Proof, not promises

Three things you can open right now

Most synthetic-data claims are unverifiable. Atlas ships the evidence.

📊

Fidelity scorecard

Every sourced module is generated as a focused cohort and checked against its cited public target (NHANES, CDC, SEER, AHA) by the cohort harness.

565 / 565metric strata within tolerance · 100/100 modules pass

Open the scorecard →

📉

SDoH is causal, not a tag

Social-risk burden reduces what actually gets generated — fewer ambulatory visits, fewer filled prescriptions. The relationship a model needs to learn.

−39% / −32%ambulatory encounters / medication fills at high burden

Open the benchmark →

🧬

A library that extends itself

atlas author turns a condition name into a citation-grounded draft module and its sourced fidelity expectation — gated by clinician sign-off.

dossier → moduleresearch · synthesize · validate · promote

How authoring works →

The differentiator, measured

Patients with barriers behave like patients with barriers

Generated with --with-sdoh over a chronic-disease cohort. As a patient's positive SDoH screens rise, utilization falls — a gradient a tag-only generator can't produce.

Positive SDoH screens	Mean ambulatory encounters	vs. none	Mean medication fills	vs. none
0 (none)	1.24	—	0.45	—
1	1.04	−16%	0.38	−17%
2	0.86	−31%	0.31	−32%
3+	0.76	−39%	0.31	−32%

Rates calibrated to BRFSS care-avoidance and Urban Institute cost-related non-adherence figures.

What it looks like

FHIR-native, profile-conformant output

One transaction Bundle per patient. Every resource targets a named US Core 6.1 profile and carries the HL7 HTEST tag so synthetic data is never mistaken for real.

Resources: Patient, Condition, Observation, Encounter, MedicationRequest, Procedure, Immunization, DiagnosticReport, Coverage, Claim/EOB, MeasureReport, DocumentReference.
Formats: FHIR R4/R5 Bundles, $export-aligned NDJSON, and Parquet.
Identity: every patient gets a Parker GPX identifier under the synthetic namespace.
Grounded notes: template or LLM-authored progress notes grounded in structured data.

{
  "resourceType": "Patient",
  "id": "GPX-SYN-0000000042-7",
  "meta": {
    "profile": ["…/us-core-patient|6.1.0"],
    "tag": [{ "code": "HTEST",
      "display": "test health data" }]
  },
  "gender": "female",
  "birthDate": "1968-04-12"
}
{
  "resourceType": "Observation",
  "meta": { "profile": ["…/us-core-blood-pressure|6.1.0"] },
  "category": [{ "coding": [{ "code": "vital-signs" }]}],
  "code": { "text": "Blood pressure panel" },
  "component": [
    { "code": { "text": "Systolic" },
      "valueQuantity": { "value": 152, "unit": "mm[Hg]" }},
    { "code": { "text": "Diastolic" },
      "valueQuantity": { "value": 96, "unit": "mm[Hg]" }}
  ]
}

Quick start

From clone to cohort in a minute

Apex Atlas is a Python package. Generate a validated FHIR cohort locally, then check it against public norms.

git clone https://github.com/ParkerApex/apex-atlas.git && cd apex-atlas
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

# Generate 1,000 synthetic FHIR R4 patients with SDoH causal modeling
atlas generate --patients 1000 --seed 42 --with-sdoh --out ./out

# Validate a cohort's prevalence against its cited public targets
atlas generate --patients 20000 --seed 42 --module hypertension --out ./cohort
atlas validate ./cohort --cohort --module hypertension

# Draft a brand-new, citation-grounded module from a research dossier
atlas author synthesize --dossier ./glaucoma.dossier.yaml --out ./atlas-drafts

Prefer to look before you install? Browse a ready-made 25-patient sample cohort → — FHIR R4 bundles with conditions, meds, grounded notes, SDOHCC screening, payer coverage, and HEDIS MeasureReports.