syntha-ehr 0.5.0

Synthetic patient record generator (Synthea-inspired) trained on pristine-healthy episode data

🩺 syntha

A Synthea-inspired hybrid synthetic patient record generator — learns the joint distribution of real anonymized Turkish-cohort EHR episodes with a Gaussian copula, then layers Synthea-style clinical pathways on top to emit fully-coded FHIR R4 bundles in Turkish.

---

🖥️ Desktop app — generate synthetic patients without code

   

_{A Tauri 2 desktop app that bundles the trained Gaussian copula and samples synthetic patients fully client-side (no Python required). Pick cohort + n + seed + constraints, hit Generate, and download a CSV.}

📦 Installers are produced by the release workflow on every v* tag push and live at stable filenames (syntha_aarch64.dmg, syntha_x64-setup.exe, syntha_amd64.AppImage). The buttons above all use releases/latest/download/… so they track the latest release automatically — no manual link maintenance per version. A daily Install-buttons verification workflow HEAD-checks each URL and opens an issue if any 404s. Source for the app lives in app/.

🛡️ macOS sees "syntha.app" is damaged? That's Gatekeeper's misleading error for unsigned apps. Until the signing pipeline ships (app/README.md → signing setup), strip the quarantine flag manually:

xattr -dr com.apple.quarantine /Applications/syntha.app

---

📑 Table of contents

• 🔍 Why syntha?
• 🎯 What it produces
• ⚠️ The catch (what it is not)
• 🇹🇷 Turkish cohort + Turkish output
• 🧪 Use cases
• 🚀 Quick start
• 📊 Distribution fidelity
• 📦 Example output
• 🌐 FHIR endpoints
• 🧱 Architecture
• 🧬 Synthea-style clinical modules
• 🛠️ CLI reference
• 🗺️ Roadmap
• 🤝 Contributing + clinician curation
• 📄 License + citation

---

🔍 Why syntha?

Synthea is the gold standard for synthetic FHIR patients, but it is rules-only and tuned to US population priors. CTGAN-style purely-generative models capture data faithfully but emit physiologically impossible tuples and have no clinical-pathway awareness. syntha gives you both:

	Synthea (rules-only)	CTGAN / copula-only	syntha (hybrid)
Matches this cohort's lab distributions	❌ generic US priors	✅	✅
Coherent prescriptions per condition	✅	❌	✅
Physiologically valid (BP, eGFR…)	✅	⚠️ sometimes	✅
LOINC + SNOMED + ICD-10 + RxNorm-coded FHIR	✅	❌	✅
Longitudinal trajectories	✅ state machines	❌	✅ drift + sticky flags
Turkish locale (names, addresses, displays)	❌	❌	✅

🎯 What it produces

For each synthetic patient, syntha emits a FHIR R4 transaction Bundle containing:

• 👤 Patient — Turkish HumanName + Address + tr language code, derived birthDate
• 🧪 Observation × ~12 — LOINC-coded labs and vitals (glucose, lipid panel, CBC, LFTs, eGFR, BP, …)
• 🩺 Condition × N — every active comorbidity flag, dual-coded SNOMED CT + ICD-10, with English/Turkish display text
• 🏥 Encounter × M — one per active condition, driven by the relevant clinical module
• 💊 MedicationRequest × P — RxNorm-coded, dosage included
• 🔬 Procedure × Q — e.g. HbA1c, lipid panel, ECG, spirometry
• 📋 CarePlan × R — disease-specific lifestyle / monitoring plans

Plus a flat CSV that matches the input schema for drop-in use as training data.

⚠️ The catch (what it is not)

• 🚫 Not a substitute for real PHI when validity hinges on rare events — the copula reproduces the bulk of the joint distribution, not the long tails.
• 🚫 Not privacy-proof. Gaussian copulas are not differentially private; if the source has fewer than ~50 patients with a rare combination, syntha may reproduce that combination too closely. Do not use when the source is a small sensitive cohort without adding a DP mechanism.
• 🚫 No disease progression simulator yet — the copula gives a cross-sectional snapshot; longitudinal mode adds plausible drift but is not a Synthea-PADM state machine. (See v0.8 in the roadmap.)
• 🚫 The source CSVs are anonymized retrospective Turkish-cohort episodes of healthy patients — synthetic disease prevalence is lower than Turkish national averages (TÜİK). If you need a population-representative Turkish cohort, calibrate per the v0.6 roadmap items.
• ⚠️ Continuous↔binary correlations are attenuated ~50% in magnitude (signs are correct since v0.3.2). Pure Spearman rank correlation on tied binary columns is biased toward zero; the proper fix is the polyserial/tetrachoric correlation, queued as v0.4 in the roadmap. For most downstream uses (training risk models, healthy-control comparisons) this is acceptable; if you need exact lab↔disease correlations, wait for v0.4 or contribute the fix.

🇹🇷 Turkish cohort + Turkish output

The training data comes from pristine_strict_episodes.csv and pristine_tolerant_episodes.csv — anonymized retrospective EHR episodes from a Turkish patient cohort selected to represent clinically pristine (i.e. healthy / minimally medicated) adults. Source CSVs are never committed to this repo (gitignored).

Synthetic output is Turkish-localized:

• Patient names sampled from common Turkish given-name and family-name distributions (src/syntha/locale/turkish.py).
• Addresses use real Turkish cities weighted by approximate population, with ISO 3166-2:TR province codes.
• Every Condition emits both an English SNOMED display and a clinical-Turkish translation in Condition.code.text.
• Patient.communication is set to tr.

All clinical terminology used (LOINC, SNOMED CT, ICD-10, RxNorm) comes from open international standards — no licensed terminology content is reproduced or embedded.

🧪 Use cases

Where to use it	Why
🤖 Training ML risk models without exposing real PHI	The copula preserves joint distributions, so a model trained on synthetic data transfers reasonably to real test sets (TSTR benchmark in v0.9).
🧬 Bioinformatics healthy-control cohorts	The source is pristine healthy episodes — use the synthetic patients as a normal-baseline group to compare against your disease cohort.
🛠️ EHR pipeline / ETL integration testing	Realistic-but-fake FHIR R4 bundles with valid LOINC/SNOMED/ICD-10/RxNorm codes are ideal for testing FHIR consumers, mapping pipelines, and OMOP/i2b2 ETLs without DPA paperwork.
📚 Teaching / coursework	Drop-in dataset for biostatistics, epidemiology, and clinical-informatics teaching without IRB.
🔬 Data augmentation	Boost rare-event coverage by oversampling synthetic patients with specific comorbidity combinations (conditional sampling lands in v0.7).

🚀 Quick start

# 1. Install
git clone https://github.com/ArioMoniri/syntha.git
cd syntha
pip install -e .

# 2. (Optional) Ingest your source CSVs — files in data/raw/ are gitignored
bash scripts/ingest_csvs.sh

# 3. Generate 1000 synthetic episodes + FHIR bundles + model card + validation report
syntha generate \
  --input data/raw/pristine_tolerant_episodes.csv \
  --output output/tolerant \
  --n 1000 --cohort tolerant

# 4. Longitudinal — 500 baseline patients × ~4 encounters over 3 years
syntha generate \
  --input data/raw/pristine_tolerant_episodes.csv \
  --output output/tolerant_long \
  --n 2000 --cohort tolerant \
  --longitudinal --encounters-per-patient 4 --years-of-history 3

# 5. Validate any synthetic CSV against its source
syntha validate \
  --source data/raw/pristine_tolerant_episodes.csv \
  --synthetic output/tolerant/synthetic_tolerant_episodes.csv \
  --output output/tolerant/validation.json

📊 Distribution fidelity

A 100-episode sample of tolerant cohort vs the full 135 569-row source:

Marginal distributions

Spearman correlation structure

Disease prevalence

Numbers (from examples/sample_output/sample_validation_report.json)

Metric	Value
n (source / synthetic)	135 569 / 100
Max Kolmogorov–Smirnov across continuous columns	0.14
Mean KS	0.07
Max binary-prevalence error	0.025 (has_rx_data)
Disease-prevalence error (HTN / DM / hyperlipidemia)	0.015 / 0.004 / 0.010
Spearman correlation-matrix Frobenius diff	2.94

📝 The KS statistic is well below the typical 0.20 "noticeable difference" threshold for every column; binary marginals (gender, disease prevalence) match to within ~1 percentage point.

📦 Example output (embedded)

A pretty-printed sample FHIR Bundle, a 100-episode synthetic CSV, the model card, and the validation report all live under examples/sample_output/ and are tracked in git.

File	Click to view (GitHub built-in viewer)	What's inside
🧾 Full FHIR Bundle (pretty)	sample_bundle_pretty.json	One transaction Bundle: Patient + Observations + Conditions + Encounter + MedicationRequests + Procedure + CarePlan
📡 100 bundles, NDJSON	sample_bundles.ndjson	Bulk-FHIR-style export, one transaction Bundle per line
📊 Flat CSV	sample_episodes.csv	100 synthetic episodes matching input schema
🗒️ Model card	sample_model_card.json	source sha256, n_train, marginals, top correlations
✅ Validation report	sample_validation_report.json	KS / Wasserstein / correlation-Frobenius per column

💡 Embedded viewer. GitHub renders the linked JSON files with syntax highlighting and a collapsible outline (click the {} icon top-right of the file view). For full FHIR-aware validation and tree-view rendering, drag the file onto simplifier.net or paste it into the official HL7 Clinical FHIR Renderer.

👁️ Inline preview — first synthetic patient (click to expand)

{
  "resourceType": "Bundle",
  "type": "transaction",
  "timestamp": "2017-05-27T21:49:42Z",
  "entry": [
    {
      "resource": {
        "resourceType": "Patient",
        "id": "20f13c43-d17b-443b-b7a7-69ccc40631c6",
        "gender": "male",
        "name": [{"use": "official", "family": "Avcı", "given": ["Furkan"]}],
        "address": [{
          "use": "home", "type": "physical",
          "city": "İstanbul", "state": "TR-34", "country": "TR"
        }],
        "communication": [{
          "language": {"coding": [{"system": "urn:ietf:bcp:47", "code": "tr", "display": "Turkish"}]},
          "preferred": true
        }],
        "birthDate": "1975-…"
      }
    },
    {
      "resource": {
        "resourceType": "Observation",
        "code": {
          "coding": [{"system": "http://loinc.org", "code": "8480-6",
                       "display": "Systolic blood pressure"}]
        },
        "valueQuantity": {"value": 118.72, "unit": "mm[Hg]"}
      }
    },
    {
      "resource": {
        "resourceType": "Condition",
        "code": {
          "coding": [
            {"system": "http://snomed.info/sct", "code": "414545008",
             "display": "Ischemic heart disease (disorder)"},
            {"system": "http://hl7.org/fhir/sid/icd-10", "code": "I25.9",
             "display": "Chronic ischaemic heart disease, unspecified"}
          ],
          "text": "Ischemic heart disease (disorder) / İskemik kalp hastalığı"
        }
      }
    },
    {
      "resource": {
        "resourceType": "MedicationRequest",
        "medicationCodeableConcept": {
          "coding": [{
            "system": "http://www.nlm.nih.gov/research/umls/rxnorm",
            "code": "243670", "display": "Aspirin 81 MG Oral Tablet"
          }]
        },
        "dosageInstruction": [{"text": "81 mg daily"}]
      }
    }
  ]
}

👁️ Inline preview — first 5 rows of the CSV

RF_EPISODE2	HASTA_ID	episode_date	gender	age	bp_sys	bp_dia	hdl	ldl	hgb	egfr	Hipertansiyon	DM_Tum
92893619	SYN_7D70431D	2017-05-27	M	42	118.7	63.0	95.0	58.0	12.9	105.7	0	0
…	…	…	…	…	…	…	…	…	…	…	…	…

Full file: examples/sample_output/sample_episodes.csv (100 rows × 73 cols).

👁️ Inline preview — validation report summary

{
  "n_source": 135569,
  "n_synthetic": 100,
  "ks_max": 0.14,
  "ks_mean": 0.07,
  "binary_max_abs_error": 0.025,
  "correlation_frobenius": 2.94
}

🌐 FHIR endpoints

syntha emits canonical FHIR R4 resources, so every emitted resource type maps to its standard REST endpoint:

Resource type	GET (read)	GET (search)	Create (POST to base)
👤 Patient	GET /Patient/{id}	GET /Patient	as part of transaction Bundle
🧪 Observation	GET /Observation/{id}	GET /Observation?subject={ref}	↑
🩺 Condition	GET /Condition/{id}	GET /Condition?patient={id}	↑
🏥 Encounter	GET /Encounter/{id}	GET /Encounter?patient={id}	↑
💊 MedicationRequest	GET /MedicationRequest/{id}	GET /MedicationRequest?patient={id}	↑
🔬 Procedure	GET /Procedure/{id}	GET /Procedure?patient={id}	↑
📋 CarePlan	GET /CarePlan/{id}	GET /CarePlan?patient={id}	↑
📦 Bundle	GET /Bundle/{id}	—	POST / (transaction)

Spin up a demo FHIR server locally

syntha serve --bundles examples/sample_output/sample_bundles.ndjson --port 8080

Then:

curl http://127.0.0.1:8080/metadata           # CapabilityStatement
curl http://127.0.0.1:8080/Patient            # searchset Bundle (all Patients)
curl http://127.0.0.1:8080/Patient/{id}       # single Patient
curl http://127.0.0.1:8080/Observation        # all Observations
curl http://127.0.0.1:8080/\$export           # FHIR Bulk Data export (NDJSON)

This is a read-only demo server (stdlib http.server, no dependencies). For a production-grade FHIR server, POST the bundles to a HAPI / Microsoft FHIR / Google Healthcare API instance — see below.

POST the bundles to any FHIR R4 server

scripts/post_to_fhir.sh POSTs every transaction Bundle in an NDJSON file to a configurable FHIR endpoint (default: the public HAPI test server):

# To the public HAPI playground:
bash scripts/post_to_fhir.sh examples/sample_output/sample_bundles.ndjson

# To your own server:
FHIR_BASE=http://localhost:8080/fhir bash scripts/post_to_fhir.sh

Once uploaded, you can browse the resources in any FHIR UI — e.g. HAPI's built-in browser or the Open Patient Browser.

🧱 Architecture

┌──────────────┐    ┌──────────────────┐    ┌──────────────────────┐
│  Source CSV  │──▶│  Gaussian copula  │──▶│ Physiologic filter   │
│ (Turkish     │    │ (Spearman → ρ;   │    │ (BP, Friedewald,     │
│  pristine)   │    │ nearest-PSD)     │    │  eGFR/creatinine)    │
└──────────────┘    └──────────────────┘    └─────────┬────────────┘
                                                       │
                                  ┌────────────────────┴────────────────────┐
                                  │                                         │
                                  ▼                                         ▼
                       ┌──────────────────┐                  ┌──────────────────────────┐
                       │ Longitudinal     │   (optional)     │  Direct single-episode   │
                       │ expansion        │ ───────────────▶│  CSV + FHIR R4 export     │
                       │ (drift, Poisson) │                  │  with Synthea-style       │
                       └─────────┬────────┘                  │  module activation        │
                                 │                            └──────────────────────────┘
                                 ▼
                          (same FHIR export)

Read docs/ARCHITECTURE.md for the math (Spearman→Gaussian transform, nearest-PSD projection, constraint rules).

🧬 Synthea-style clinical modules

Nine modules ship out of the box (src/syntha/modules/); each fires on its corresponding source-CSV comorbidity flag:

Module	Flag(s)	Emits
🫀 Hypertension	Hipertansiyon	Encounter, 1–2 antihypertensives (stage 2 → dual), CarePlan
🍬 Diabetes	DM_Tum, DM_Komplikasyonlu	Encounter, HbA1c, metformin (+ insulin if severe), CarePlan
🧀 Hyperlipidemia	Hiperlipidemi	Encounter, lipid panel, statin (high-intensity if LDL ≥ 190)
🦋 Thyroid	Tiroid	Encounter, TSH, levothyroxine
😔 Depression	Depresyon	Psych encounter, sertraline, CBT CarePlan
😰 Anxiety	Anksiyete	Psych encounter, escitalopram (or buspirone if already on SSRI)
❤️ IHD	Iskemik_Kalp	Cardiology encounter, ECG, aspirin + β-blocker + statin
🌬️ Asthma	Astim	Resp encounter, spirometry, SABA + ICS
🚭 COPD	COPD	Resp encounter, spirometry, LABA + SABA

See docs/MODULES.md for the authoring guide. Clinician contributions for TR-specific drug choices are highly welcome — see CONTRIBUTING.md.

🛠️ CLI reference

Command	Description
syntha generate	End-to-end: train copula + sample + modules + CSV/FHIR + model card + validation report
syntha fit	Fit and persist a copula in a registry without sampling
syntha sample	Raw sampling from a registered model
syntha fhir	Convert an existing synthetic CSV to FHIR bundles
syntha validate	KS / Wasserstein / correlation diff between source and synthetic
syntha serve	Boot a read-only FHIR R4 demo server from a bundles NDJSON file
syntha export-model	Export a registered copula to a compact JSON the desktop app consumes
syntha list-models	List models in a registry
syntha show-card	Print a model card

Run syntha <cmd> --help for full option lists.

🗺️ Roadmap

The full phased roadmap (v0.1 → v1.0) lives in ROADMAP.md. Highlights:

• v0.6 — clinician curation 🟣 — needs Dr. Moniri (or a collaborator)
• v0.7 — optional CTGAN/TVAE backend ⬜
• v0.8 — true Synthea PADM-style state machines ⬜
• v0.9 — TSTR benchmark ⬜
• v1.0 — PyPI + paper ⬜

🤝 Contributing + clinician curation welcome

There are three ways to feed clinical guidance into syntha — pick whichever is least friction for you:

1. 🚀 Just tell me (lowest friction)

Reply in any open conversation with Claude (the agent that maintains this repo) saying e.g.

"In Türkiye, perindopril 5 mg is the typical first-line ACEi for uncomplicated hypertension per TKD 2023 — switch the default in the hypertension module."

…and I'll edit the relevant file, push, and re-run CI. No GitHub UI needed.

2. 📝 GitHub issue (recommended for asynchronous tracking)

Open an issue using the 🧑‍⚕️ Clinical curation template — one click:

👉 Open a Clinical curation issue 👈

The template pre-lists the files most likely to need changes:

If you want to change…	Edit this file
Which drug a module prescribes	src/syntha/modules/<condition>.py
The RxNorm code or dose text	src/syntha/fhir/rxnorm.py
The SNOMED / ICD-10 code for a Condition	src/syntha/fhir/codes.py
Turkish display strings	src/syntha/locale/turkish.py
Prevalence calibration / disease-progression rules	src/syntha/longitudinal.py

3. 🔧 Pull request

git clone https://github.com/ArioMoniri/syntha
cd syntha
pip install -e ".[dev]"
# … edit files …
pytest -q
git checkout -b clinical/<short-topic>
git commit -am "clinical: <what you changed and why>"
git push -u origin clinical/<short-topic>
gh pr create   # or open via the GitHub UI

What's currently flagged 🟣 (waiting for clinician input)

Per ROADMAP.md → v0.6:

• 🟣 TR-specific first-line drug calibration — current defaults are international (lisinopril/amlodipine for HTN, metformin for DM, atorvastatin for hyperlipidemia). Turkish primary-care reality may differ (e.g. perindopril, nebivolol).
• 🟣 New modules: CKD staging (eGFR-driven), MAFLD (ALT/AST + obesity), anemia (Hb-driven), B12 deficiency (vit B12 column directly available).
• 🟣 Prevalence calibration to TÜİK — synthetic disease rates currently mirror the pristine-healthy source cohort. To use syntha as a Turkish-population baseline rather than a healthy baseline, the marginals should be calibrated to TÜİK figures.
• 🟣 Turkish display string review — confirm clinical-Turkish preferred terms match Türk Tabipleri Birliği / TR-specific usage rather than literal translations.
• 🟣 ICD-10 specificity — the current mapping uses unspecified (".9") forms; specifying further (E11.65, I50.32, etc.) when the source flag carries the information would improve downstream realism.

Full developer guide: CONTRIBUTING.md. All PRs must pass the CI matrix (Py 3.10 → 3.13) before merge.

📄 License + citation

Apache 2.0 © 2026 Ariorad Moniri — see LICENSE.

If you use syntha in academic work, please cite:

Moniri, A. (2026). syntha: hybrid synthetic patient record generator
trained on Turkish pristine-healthy EHR cohorts.
https://github.com/ArioMoniri/syntha

---

Acknowledgements

• 🩺 Synthea — the inspiration for the clinical-module layer and FHIR output format.
• 🌐 Open clinical terminologies: LOINC, SNOMED CT, ICD-10, RxNorm.
• 📊 The anonymized Turkish-cohort EHR data used to train the copula (de-identified by the upstream data steward; never redistributed by this repo).

Contributors

Ariorad Moniri 💻 🎨 📖 🚧 🤔 👀 🚇 ⚠️

This project follows the all-contributors specification — contributions of any kind welcome. Comment @all-contributors please add @username for code,doc on an issue or PR to nominate someone.

💬 Community

• 🗨️ GitHub Discussions — open questions, "is this the right tool for X?", show-and-tell
• 🐛 Issues — bug reports + feature requests + clinical-curation
• 📖 Contributing — dev setup + commit conventions + clinical-curation workflow