xpertsystems/hccar001-sample

HCCAR001 — Synthetic Heart Failure Dataset (Sample Preview)

XpertSystems.ai | Synthetic Data Factory | Healthcare / Cardiology Vertical

A two-table longitudinal heart failure patient dataset spanning the full clinical-research data surface for HF cohorts: baseline patient records (~114 features per patient covering demographics, full echocardiographic assessment with diastolic function, 14 biomarkers including BNP/NT-proBNP/troponin/CKD-EPI eGFR, guideline-directed medical therapy (GDMT) 4-pillar prescribing, device therapy (ICD/CRT/ LVAD/transplant), hospitalization outcomes, functional status (NYHA, 6MWD, KCCQ, VO₂ max), comorbidities, and vital signs) plus quarterly follow-up visits over 3 years tracking LVEF/BNP/NYHA trajectories.

Calibrated benchmark-first against ACC/AHA 2022 Heart Failure Guidelines, ASE 2016 Recommendations for Evaluation of LV Diastolic Function, CKD-EPI 2009, PINNACLE Registry (real-world GDMT prescribing rates), CMS HRRP (30-day readmission benchmarks), SCD-HeFT / DINAMIT (primary-prevention ICD criteria), and CARE-HF / COMPANION (CRT eligibility criteria).

This is the sample preview — 200 patients × 12 quarterly visits over 3 years (200 baseline records + 2,400 visit records, ~500 KB). The full product covers 10,000+ patients × full 3-year follow-up with extended echocardiographic / CMR modules, full CPET / device interrogation detail, complete medication titration trajectories, and pre-built cohort configs for HFrEF clinical trial simulation, HFpEF treatment-effect heterogeneity studies, and cardiorenal-cardiometabolic comorbidity analysis.

---

Dataset summary

Table	Rows (sample)	What it contains
baseline	200	One row per patient. 114 features spanning: demographics + enrollment, cardiac function (LVEF, LV dimensions, GLS, RV function, TAPSE), diastolic function (E/A, E/e', LAVI, TRPG, diastolic grade), biomarkers (BNP, NT-proBNP, troponin I/T, creatinine, eGFR, BUN, electrolytes, hemoglobin, CRP, albumin, uric acid, iron studies, CKD stage), GDMT 4-pillar + ARNI/ivabradine/hydralazine-nitrate + device therapy (ICD/CRT/LVAD/transplant), hospitalization (LOS, ICU, 30/90/180-day readmission, ED visits, hospitalization risk score), functional outcomes (6MWD, KCCQ, VO₂ max, EF response, NYHA improvement, MACE, mortality, time-to-event), comorbidities (AF, diabetes, HTN, CAD, valvular, sleep apnea, COPD, anemia, obesity, depression, CCI), and vital signs (HR, SBP/DBP, MAP, weight, SpO₂, JVP, edema grade, rales)
visits	2,400	Quarterly visit-level records over 12 visits × 3 years. LVEF trajectory, BNP trajectory, NYHA class updates, weight changes — useful for longitudinal modeling, treatment-response trajectories, and time-to-event ML

Both tables provided in CSV and Parquet. Join on patient_id.

---

Calibration sources

All ten validation metrics target named clinical / regulatory standards:

• ACC/AHA 2022 HF Guidelines — HF phenotype LVEF bands (HFrEF ≤40%, HFmrEF 41-49%, HFpEF ≥50%, Advanced <30%)
• ASE 2016 Recommendations for Evaluation of LV Diastolic Function — E/A ratio, E/e', LAVI, TRPG diagnostic criteria
• Lang et al. (ASE 2015) — chamber quantification recommendations for LVEDV/LVESV/SV identities
• CKD-EPI 2009 — eGFR formula physiologic bounds [8, 140]
• PINNACLE Registry — real-world GDMT 4-pillar prescribing rates for outpatient HFrEF cohorts
• CMS HRRP (Hospital Readmissions Reduction Program) — 30-day HF readmission national benchmark 22-25% + temporal monotonicity
• SCD-HeFT (Bardy et al. 2005) + DINAMIT (Hohnloser et al. 2004) — primary prevention ICD eligibility (LVEF ≤35%, NYHA II-III)
• CARE-HF (Cleland et al. 2005) + COMPANION (Bristow et al. 2004) — CRT eligibility (LVEF ≤35%, NYHA III-IV ambulatory, LBBB)

---

Validation scorecard (seed = 42)

10/10 PASS · Grade A+ (100%) across all six canonical seeds (42, 7, 123, 2024, 99, 1).

#	Metric	Observed	Target	Tol	Type	Source
1	lvef_in_phenotype_band_rate	1.000	0.99	±0.01	FLOOR	ACC/AHA 2022
2	lvesv_equals_lvedv_times_one_minus_ef_rate	1.000	0.99	±0.01	FLOOR	Lang et al. ASE 2015
3	stroke_volume_equals_lvedv_minus_lvesv_rate	1.000	0.99	±0.01	FLOOR	ASE 2015
4	cardiac_index_equals_co_over_bsa_rate	1.000	0.99	±0.01	FLOOR	Hemodynamic identity
5	egfr_in_physiologic_bounds_rate	1.000	0.99	±0.01	FLOOR	CKD-EPI 2009
6	gdmt_4_pillar_requires_lvef_under_40_rate	1.000	0.99	±0.01	FLOOR	ACC/AHA 2022 + PINNACLE
7	icd_requires_lvef_under_35_and_nyha_ge_2_rate	1.000	0.99	±0.01	FLOOR	SCD-HeFT / DINAMIT
8	crt_requires_lvef_under_35_and_nyha_ge_3_rate	1.000	0.99	±0.01	FLOOR	CARE-HF / COMPANION
9	readmission_temporal_ordering_rate	1.000	0.99	±0.01	FLOOR	CMS HRRP
10	ea_ratio_equals_e_over_a_rate	1.000	0.99	±0.01	FLOOR	ASE 2016

---

Schema highlights

baseline (200 rows × 114 cols)

Demographics (11 cols): patient_id (UUID), site_id, hf_phenotype (HFrEF / HFmrEF / HFpEF / Advanced), age_at_baseline, sex_male, race_ethnicity (5 categories), bmi_kg_m2, bsa_m2, nyha_class_baseline (1-4), index_hospitalization, enrollment_date.

Cardiac function (13 cols): lvef_pct_baseline, lvedd_mm_baseline, lvesd_mm_baseline, lvedv_ml_baseline, lvesv_ml_baseline, stroke_volume_ml_baseline, heart_rate_baseline_bpm, cardiac_output_l_min_baseline, cardiac_index_l_min_m2_baseline, gls_pct_baseline, rv_function_baseline, tapse_mm_baseline, echo_quality_baseline.

Diastolic function (9 cols): diastolic_grade_baseline (Grade_1 / Grade_2 / Grade_3 / Indeterminate), e_velocity_cm_s_baseline, a_velocity_cm_s_baseline, ea_ratio_baseline, e_prime_cm_s_baseline, e_e_prime_ratio_baseline, lavi_ml_m2_baseline, trpg_mmhg_baseline, ivrt_ms_baseline.

Biomarkers (15 cols): bnp_pg_ml_baseline, nt_probnp_pg_ml_baseline, troponin_i_ng_ml_baseline, troponin_t_ng_ml_baseline, creatinine_mg_dl_baseline, egfr_ml_min_173m2_baseline, bun_mg_dl_baseline, sodium_meq_l_baseline, potassium_meq_l_baseline, hemoglobin_g_dl_baseline, iron_studies_baseline, crp_mg_l_baseline, albumin_g_dl_baseline, uric_acid_mg_dl_baseline, ckd_stage_baseline.

Treatment & GDMT (14 cols): gdmt_acei_arb_arni, gdmt_betablocker, gdmt_mra, gdmt_sglt2i, gdmt_4_pillar_flag, arni_flag, ivabradine_flag, hydralazine_nitrate_flag, icd_flag, crt_flag, lvad_flag, transplant_flag, diuretic_loop_dose_mg, gdmt_dose_optimization_pct.

Hospitalization (13 cols): hospitalization_count_3yr, index_los_days, index_icu_flag, readmission_30d_flag, readmission_90d_flag, readmission_180d_flag, readmission_cause, iv_diuretic_flag, diuretic_response_ml_mg, discharge_congestion_score, ed_visit_count_1yr, days_to_readmission, hospitalization_risk_score.

Functional outcomes (15 cols): six_mwd_meters_baseline, kccq_overall_score_baseline, vo2_max_ml_kg_min_baseline, lvef_change_12m_pct, ef_response_flag, nyha_improvement_flag, six_mwd_trend_12m_meters, kccq_change_12m, mace_flag_3yr, mortality_flag_3yr, cv_death_flag_3yr, time_to_death_days, borg_dyspnea_scale_baseline, patient_global_impression_12m, cpet_reason.

Comorbidities & vitals (25 cols): atrial_fibrillation_flag, af_type, diabetes_flag, diabetes_type, hypertension_flag, cad_flag, prior_mi_flag, valvular_disease, sleep_apnea_flag, copd_flag, anemia_flag, obesity_flag, depression_flag, charlson_comorbidity_index, heart_rate_bpm_baseline, systolic_bp_mmhg_baseline, diastolic_bp_mmhg_baseline, pulse_pressure_mmhg_baseline, map_mmhg_baseline, weight_kg_baseline, oxygen_saturation_pct_baseline, jvp_cmh2o_baseline, peripheral_edema_grade_baseline, rales_flag_baseline.

visits (2,400 rows × 11 cols)

patient_id, visit_number (1-12), visit_date, months_from_baseline (0-33), hf_phenotype, nyha_class_visit, lvef_pct_visit, lvef_trend_visit, bnp_pg_ml_visit, bnp_trend_pct_visit, weight_change_kg_visit.

---

Suggested use cases

• HF phenotype classification ML — train classifiers for HFrEF / HFmrEF / HFpEF / Advanced from echo + biomarker features
• 30-day readmission prediction — classifier on readmission_30d_flag from demographics, biomarkers, GDMT status, hospitalization risk features (CMS HRRP-relevant)
• GDMT optimization ML — predict which patients are eligible for 4-pillar therapy, which are likely to achieve target doses
• Mortality / MACE risk stratification — Cox models or survival forests on mortality_flag_3yr + time_to_death_days with right- censoring
• EF response prediction — regressor for lvef_change_12m_pct from baseline echo + GDMT + biomarkers; identifies likely responders
• Diastolic function classification — train multi-class on diastolic_grade_baseline (Grade 1/2/3/Indeterminate) from E/A, E/e', LAVI features per ASE 2016
• Cardiorenal risk modeling — joint modeling of egfr_ml_min_173m2 trajectory and bnp_pg_ml trajectory across visits; useful for cardiorenal syndrome detection
• NLP-augmented EHR research — table provides structured ground truth for NLP extraction from clinical notes; pair with synthetic clinical text generators
• Clinical trial cohort simulation — filter to specific eligibility criteria (e.g., HFrEF + NYHA II-III + LVEF ≤35% + eGFR ≥30) and benchmark expected event rates / power calculations
• Treatment effect heterogeneity — train uplift / CATE models for SGLT2i, ARNI, MRA effects across phenotype subgroups
• Longitudinal LVEF / BNP trajectory clustering — unsupervised clustering on visit-level trajectories to discover responder phenotypes
• NYHA class transitions — Markov / state-space models on nyha_class_visit transitions
• Device therapy appropriateness ML — train models that flag potential under-utilization of ICD/CRT in eligible patients
• Hospitalization cost / utilization modeling — combine index_los_days, index_icu_flag, ed_visit_count_1yr, hospitalization_count_3yr for utilization predictive analytics

---

Loading examples

from datasets import load_dataset

baseline = load_dataset("xpertsystems/hccar001-sample", "baseline", split="train")
visits = load_dataset("xpertsystems/hccar001-sample", "visits", split="train")
print(baseline.shape, visits.shape)

import pandas as pd
from huggingface_hub import hf_hub_download

baseline = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar001-sample", "hccar001_baseline.parquet",
    repo_type="dataset",
))
visits = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar001-sample", "hccar001_visits.parquet",
    repo_type="dataset",
))

# HF phenotype distribution (ACC/AHA 2022)
print(baseline["hf_phenotype"].value_counts(normalize=True).round(3))

# GDMT 4-pillar prescribing in HFrEF
import pandas as pd
from huggingface_hub import hf_hub_download

baseline = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar001-sample", "hccar001_baseline.parquet",
    repo_type="dataset",
))

hfref = baseline[baseline["hf_phenotype"] == "HFrEF"]
gdmt_rates = hfref[[
    "gdmt_acei_arb_arni", "gdmt_betablocker",
    "gdmt_mra", "gdmt_sglt2i", "gdmt_4_pillar_flag"
]].mean().round(3) * 100
print("HFrEF GDMT prescribing rates (%):")
print(gdmt_rates)

# Longitudinal LVEF trajectory by GDMT status
import pandas as pd
from huggingface_hub import hf_hub_download

baseline = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar001-sample", "hccar001_baseline.parquet",
    repo_type="dataset",
))
visits = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar001-sample", "hccar001_visits.parquet",
    repo_type="dataset",
))

joined = visits.merge(
    baseline[["patient_id", "gdmt_4_pillar_flag", "gdmt_sglt2i"]],
    on="patient_id"
)
joined["responder_group"] = joined["gdmt_4_pillar_flag"] | joined["gdmt_sglt2i"]
trajectory = (
    joined.groupby(["responder_group", "visit_number"])
    ["lvef_pct_visit"].mean().unstack(level=0).round(2)
)
print(trajectory)  # responder group LVEF trajectory recovers; control doesn't

# 30-day readmission risk by HF phenotype
import pandas as pd
from huggingface_hub import hf_hub_download

baseline = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar001-sample", "hccar001_baseline.parquet",
    repo_type="dataset",
))

risk = baseline.groupby("hf_phenotype").agg(
    n=("patient_id", "count"),
    readmit_30d_pct=("readmission_30d_flag", lambda x: x.mean() * 100),
    readmit_90d_pct=("readmission_90d_flag", lambda x: x.mean() * 100),
    mean_nyha=("nyha_class_baseline", "mean"),
    mean_bnp=("bnp_pg_ml_baseline", "mean"),
).round(2)
print(risk)

---

Limitations and honest disclosures

This sample is calibrated for structural fidelity, not bit-exact reproduction of any specific HF registry or institutional cohort. Specifically:

• 30-day readmission rate observed ~30-35% vs CMS HRRP national benchmark 22-25%. The generator's p_30d = 0.23 × (1 + risk × 0.8) risk-multiplier amplifies above the base rate. Use the TEMPORAL ORDERING (30d ≤ 90d ≤ 180d) as the structural validation, not the absolute rate. For absolute-rate calibration, scale to the full product or re-fit the risk multiplier.
• 3-year mortality observed ~10% vs real-world HFrEF literature 25-35% (PARADIGM-HF, EMPEROR-Reduced, DAPA-HF placebo arms). The generator computes mort_p = mace_p × 0.45 with mace_p clipped to [0.03, 0.65], producing a mort_p ceiling of ~0.29. The generator under-predicts mortality for advanced HF.
• bsa_m2 has a dead-code bug in the generator (line 120 computes BSA via Mosteller formula, line 121 immediately overwrites with N(1.9, 0.25)). The overwritten N(1.9, 0.25) is what's in the output. Don't use bsa for BMI/height back-calculation.
• LBBB is not modeled for CRT eligibility — CARE-HF / COMPANION criteria include LVEF ≤35% + NYHA III-IV ambulatory + LBBB. Our validation covers the first two (LVEF + NYHA); LBBB status is absent. Add this if you need full CRT-D appropriateness classification.
• Visit-level LVEF can drift OUT of baseline phenotype band over 12 quarterly visits as patients respond to GDMT — this is realistic clinical behavior (recovered EF). Our validation checks BASELINE LVEF in phenotype band, not visit-level. Don't filter visits by lvef_pct_visit matching baseline phenotype band — you'll lose GDMT responders.
• site_id is randomly chosen from 30 generated UUIDs — patients within a site share no clustering structure (no hierarchical / random-effects), no provider effects, no geographic correlations. Treat as a nominal site label, not a hierarchical clustering variable.
• HFpEF treatment effect modeling is simplified. GDMT prescribing rates from PINNACLE are anchored to outpatient HFrEF cohorts; HFpEF GDMT rates (acei_arb_arni: 0.55, betablocker: 0.60, mra: 0.25, sglt2i: 0.45) are reasonable but the EF response to SGLT2i in HFpEF is muted (EMPEROR- Preserved-style effects, not EMPEROR-Reduced). Our generator applies a uniform +1.5 ± 1 EF delta for SGLT2i regardless of phenotype.
• Race / ethnicity distribution (60% Non-Hispanic White, 20% Black/AA, 12% Hispanic, 5% Asian/Pacific Islander, 3% Other) matches general US prevalence but Black/AA HF prevalence is ~2x higher than this in real registries. Adjust sampling if modeling race-specific GDMT response (e.g., V-HeFT for hydralazine-nitrate).
• Hydralazine-nitrate flag fires at 20% of Black/AA patients (line 389); the generator has no separate eligibility model (ISDN should be ACEi-intolerant or sub-target dose). Treat as an inclusive eligibility signal.
• Comorbidities are sampled independently (lines 621-639) — no realistic co-occurrence structure beyond per-phenotype base rates. Real diabetes + obesity + sleep apnea + CKD exhibit strong co-occurrence (cardiometabolic phenotype); this generator treats them as independent.
• scipy.stats, faker, tqdm are mentioned in the generator's docstring but not used in active code. No external dependencies beyond NumPy + Pandas.
• NYHA class distribution by phenotype matches the design dict (HFrEF [0.10, 0.35, 0.40, 0.15]) within ~5% sampling noise at n=200. For tight matching, use the full product (10K patients).
• Visit count is fixed at 12 quarterly visits — no realistic censoring (loss to follow-up, death before visit, etc). All patients have all 12 visits regardless of mortality_flag_3yr or time_to_death_days. For survival ML, use baseline TTE variables and don't naively use visit count as risk signal.

The full HCCAR001 product addresses these by calibrated mortality and readmission rates against real-world HFrEF/HFpEF registries, hierarchical site / provider structure, dependent comorbidity sampling, full LBBB + QRS duration + bundle-branch block modeling for CRT eligibility, race- specific GDMT response heterogeneity, realistic censoring of visit records by death/LTFU, and pre-built scenario configs (HFrEF clinical trial simulation, HFpEF treatment-effect heterogeneity, cardiometabolic phenotype). Contact us for the licensed commercial release.

---

Companion datasets

This is the first SKU in our Healthcare / Cardiology vertical. Related datasets from elsewhere in the catalog:

• Healthcare / Neurology (10 SKUs) — synthetic neurological patient datasets covering stroke, MS, epilepsy, Parkinson's, ALS, traumatic brain injury, dementia spectrum
• Insurance & Risk (10 SKUs) — health insurance claims, prior authorization, risk adjustment, MLR modeling
• Energy & Climate (8 SKUs) — power grid, renewables, demand, O&G, smart grid, energy trading, climate impact, consumer electricity
• Manufacturing (10 SKUs) — reliability, quality, operations, supply chain, IIoT
• Oil & Gas (17 SKUs) — upstream, midstream, downstream, geological / seismic, reservoir simulation

For the broader catalog, see https://huggingface.co/xpertsystems

---

Citation

@dataset{xpertsystems_hccar001_sample_2026,
  author       = {XpertSystems.ai},
  title        = {HCCAR001 Synthetic Heart Failure Dataset (Sample Preview)},
  year         = 2026,
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/xpertsystems/hccar001-sample}
}

---

Contact

• Web: https://xpertsystems.ai
• Email: pradeep@xpertsystems.ai
• Full product catalog: Cardiology, Neurology, Insurance & Risk, Energy & Climate, Manufacturing, Oil & Gas, Cybersecurity, and more

Sample License: CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0) Full product License: Commercial — please contact for pricing.

Important medical disclaimer: This dataset contains SYNTHETIC patient records only. No data was derived from any real patient, EHR archive, or clinical registry. The dataset is intended for ML model development, benchmarking, and education — NOT for clinical decision support, patient counseling, or medical research conclusions. All clinical thresholds (GDMT eligibility, ICD/CRT criteria, biomarker ranges) are sourced from published guidelines; users are responsible for verifying against current ACC/AHA/HFSA guidelines for clinical applications.