vickymuller-md/heartland-synthetic

HEARTLAND Synthetic Heart-Failure Benchmark Cohort

An open, reproducible tabular benchmark containing 1,000 entirely synthetic adult heart-failure records and 31 variables. It combines clinical variables with modeled rural access, social support, guideline-directed medical therapy (GDMT), the proposed HEARTLAND risk score, and modeled one-year outcomes.

No row represents a real person. The dataset contains no protected health information (PHI), no real medical record, and no geographically accurate county identifier. It is not clinical evidence and must not be used for diagnosis, prognosis, treatment, or patient care.

Dataset summary

Property	Value
Rows	1,000
Columns	31
Format	CSV
Dataset release	1.0.0
Generator archive	heartland-synthetic v0.2.1
Current Python package	heartland-synthetic v0.2.2
Fixed seed	42
Modeled rural fraction	70%
Age range	45-95 years
HF phenotype target mix	45% HFrEF / 15% HFmrEF / 40% HFpEF
Missing values	0
Duplicate rows	0
License	MIT
SHA-256	8fbce909272274129f94db2b80b179b493e64fbc382a88515a851089b2edda8e

Links and provenance

• Dataset landing page
• Current PyPI package (v0.2.2)
• Source repository
• Generator archive used for this CSV (v0.2.1)
• Current software archive (v0.2.2)
• Software concept DOI
• HEARTLAND Protocol

The Zenodo identifiers above identify the software generator archives, not a DOI minted specifically for this CSV. Version 0.2.2 corrected packaging and publication metadata; the benchmark remains traceable to the v0.2.1 generator archive stated on its original dataset landing page.

Intended uses

Appropriate uses include:

• reproducible simulation and teaching;
• software, data-pipeline, REDCap, and FHIR interoperability testing;
• exploratory methods development with synthetic inputs;
• demonstrations of rural-access and social-support variables in heart-failure workflows;
• prototyping analyses before testing them on appropriately governed real-world data.

Inappropriate uses include:

• clinical decision-making or patient-specific risk estimation;
• estimating real population prevalence, treatment effects, disparities, or outcomes;
• replacing external validation, registry analysis, EHR analysis, or prospective research;
• using synthetic county codes as real geography;
• training or evaluating a clinical model and presenting performance as evidence of safety, effectiveness, transportability, or validation.

Loading the dataset

from datasets import load_dataset

dataset = load_dataset("vickymuller-md/heartland-synthetic")
df = dataset["train"].to_pandas()

Direct CSV loading:

import pandas as pd

url = (
    "https://huggingface.co/datasets/vickymuller-md/heartland-synthetic/"
    "resolve/main/heartland-synthetic-cohort-1000-seed42.csv"
)
df = pd.read_csv(url, dtype={"county_fips": "string"})

Regenerate cohorts with the Python package:

pip install heartland-synthetic==0.2.2

from heartland_synthetic import HeartlandCohortConfig, generate_cohort

config = HeartlandCohortConfig(
    n_patients=1000,
    rural_fraction=0.7,
    hf_type_distribution={"hfref": 0.45, "hfmref": 0.15, "hfpef": 0.40},
    age_range=(45, 95),
    include_outcomes=True,
    seed=42,
)
cohort = generate_cohort(config)

Data structure

Identifiers and modeled geography

Field	Type	Description
patient_id	string	Synthetic identifier in HS-000001 format.
state	string	Modeled US postal abbreviation. Not a real patient location.
county_fips	string	Synthetic five-character code. Not linked to a real patient or necessarily geo-accurate.
rural_urban_code	integer	Modeled RUCA-like category from 1 to 10.
distance_to_cardiology_mi	float	Modeled distance to cardiology care in miles.

Demographics and heart-failure phenotype

Field	Type	Description
age	integer	Modeled age in years.
sex	string	F or M in the current generator.
race	string	Modeled category: White, Black, Hispanic, or Other.
hf_type	string	hfref, hfmref, or hfpef.
lvef	float	Left ventricular ejection fraction, percent.

Clinical variables and comorbidities

Field	Type	Description
egfr	float	Estimated glomerular filtration rate, mL/min/1.73 m2.
bnp	float	B-type natriuretic peptide, pg/mL.
sbp, dbp	integer	Systolic and diastolic blood pressure, mmHg.
hr	integer	Heart rate, beats/minute.
bmi	float	Body mass index, kg/m2.
diabetes, af	integer	Modeled binary indicators (0/1).
ckd_stage	integer	Simplified CKD stage from 1 to 5.
ckm_stage	integer	Simplified CKM stage from 0 to 4.
social_support_score	integer	Modeled ENRICHD ESSI score from 8 to 40.
prior_hf_hosp_6mo	integer	Modeled prior HF hospitalization indicator (0/1).

Therapy, score, and modeled outcomes

Field	Type	Description
on_acei_arb_arni	integer	Modeled ACEi/ARB/ARNI exposure indicator (0/1).
on_beta_blocker	integer	Modeled beta-blocker exposure indicator (0/1).
on_mra	integer	Modeled mineralocorticoid receptor antagonist exposure indicator (0/1).
on_sglt2i	integer	Modeled SGLT2 inhibitor exposure indicator (0/1).
gdmt_classes_count	integer	Count of modeled GDMT classes from 0 to 4.
heartland_risk_score	integer	Proposed HEARTLAND score from 0 to 18.
heartland_risk_tier	string	low, moderate, or high.
mortality_1yr	integer	Modeled one-year mortality event (0/1), not an observed outcome.
hospitalization_1yr	integer	Modeled one-year hospitalization event (0/1), not an observed outcome.

Observed composition of this synthetic release

• HF phenotype: 475 HFrEF, 128 HFmrEF, and 397 HFpEF records.
• HEARTLAND tier: 361 low, 442 moderate, and 197 high records.
• Sex field: 484 F and 516 M records.
• Modeled outcomes: 170 mortality and 329 hospitalization events at one year.
• Mean age: 71.47 years; mean HEARTLAND score: 5.68.

These values describe this deterministic synthetic file only. They are not estimates of any patient population.

Generation process and source assumptions

The file was generated deterministically with seed 42. The generator combines literature-anchored marginal assumptions with correlated sampling for continuous clinical variables, rural/urban-specific assumptions for distance and social support, modeled GDMT exposure, the proposed HEARTLAND scoring rules, and tier-indexed modeled outcomes. Detailed source attributions and numerical assumptions are documented in the software README and source registries.

The source literature informs simulation assumptions; it does not validate the generated rows, the proposed score, or the modeled outcomes.

Biases, ethical considerations, and limitations

• All data are simulated. Apparent associations are produced by generator assumptions and should not be interpreted as discovered clinical relationships.
• Modeled demographic categories are simplified and do not capture the diversity, self-identification, intersectionality, or social context of real populations.
• The rurality, state, and county fields are not suitable for geographic inference or local resource allocation.
• County FIPS values are synthetic; they must not be joined to real county-level datasets as if they represented actual locations.
• Outcome rates are literature-consistent modeling assumptions, not observed events and not prospective validation.
• The HEARTLAND risk score is a proposed, unvalidated framework. This release provides no evidence of discrimination, calibration, clinical utility, fairness, transportability, or patient benefit.
• CKM staging is simplified; GDMT is modeled as point-prevalence exposure; the dataset omits many clinically relevant features.
• The dataset may be useful for software and methods testing, but good performance on synthetic data can be misleading and must never be presented as real-world clinical performance.

Integrity and versioning

File: heartland-synthetic-cohort-1000-seed42.csv

SHA-256  8fbce909272274129f94db2b80b179b493e64fbc382a88515a851089b2edda8e
Bytes    110485
Rows     1000
Columns  31

The repository history records future changes. Users should cite the exact Hugging Face revision and file hash used in an analysis.

Citation

For the dataset repository:

@dataset{muller_ferreira_heartland_synthetic_cohort_2026,
  author    = {Muller Ferreira, Vicky},
  title     = {HEARTLAND Synthetic Heart-Failure Benchmark Cohort},
  year      = {2026},
  version   = {1.0.0},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/vickymuller-md/heartland-synthetic},
  note      = {Entirely synthetic benchmark; no real patient data or PHI}
}

Also cite the archived software generator:

@software{muller_ferreira_heartland_synthetic_2026,
  author    = {Muller Ferreira, Vicky},
  title     = {heartland-synthetic: Synthetic heart-failure cohort generator with HEARTLAND risk variables},
  year      = {2026},
  version   = {0.2.1},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19635043},
  url       = {https://doi.org/10.5281/zenodo.19635043}
}

Creator and license

Created by Vicky Muller Ferreira, MD, independent researcher, HEARTLAND Protocol. ORCID: 0009-0009-1099-5690.

Released under the MIT License. See LICENSE in this repository.

Responsible-use statement

This public release supports research, education, reproducibility, and software/interoperability testing. Availability on Hugging Face does not constitute independent review, endorsement, clinical validation, regulatory clearance, adoption, or evidence of patient benefit.