Hugging Face's logo

Datasets:
xpertsystems
/
hccar003-sample

Tasks:
Tabular Classification
Tabular Regression
Time Series Forecasting
Modalities:
Tabular
Text
Formats:
parquet
Size:
1K - 10K
Tags:
synthetic-data
healthcare
cardiology
hypertension
htn
high-blood-pressure
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
hccar003-sample / README.md
pradeep-xpert's picture
pradeep-xpert
Upload folder using huggingface_hub
b440054 verified
preview code
|
raw
history blame contribute delete
22.6 kB
metadata 
license: cc-by-nc-4.0
task_categories:
  - tabular-classification
  - tabular-regression
  - time-series-forecasting
tags:
  - synthetic-data
  - healthcare
  - cardiology
  - hypertension
  - htn
  - high-blood-pressure
  - blood-pressure-monitoring
  - abpm
  - ambulatory-bp-monitoring
  - home-bp-monitoring
  - central-aortic-pressure
  - pulse-wave-velocity
  - pwv
  - augmentation-index
  - arterial-stiffness
  - bp-variability
  - white-coat-hypertension
  - masked-hypertension
  - resistant-hypertension
  - nocturnal-dipping
  - non-dipper
  - reverse-dipper
  - ace-inhibitor
  - arb
  - calcium-channel-blocker
  - ccb
  - thiazide
  - beta-blocker
  - mra
  - spironolactone
  - antihypertensive
  - medication-adherence
  - bp-response
  - side-effects
  - pill-burden
  - lifestyle-modification
  - dash-diet
  - sodium-intake
  - physical-activity
  - mets
  - sleep-quality
  - osa
  - obstructive-sleep-apnea
  - ascvd
  - ascvd-pooled-cohort
  - framingham-risk
  - pooled-cohort-equation
  - mace
  - major-adverse-cardiovascular-event
  - mi-prediction
  - stroke-prediction
  - hf-hospitalization
  - atrial-fibrillation
  - cv-death
  - lvh
  - left-ventricular-hypertrophy
  - lv-mass-index
  - e-e-prime
  - diastolic-dysfunction
  - carotid-imt
  - carotid-plaque
  - retinopathy
  - microalbuminuria
  - macroalbuminuria
  - uacr
  - ckd
  - kdigo
  - egfr
  - ckd-epi
  - ckd-stage
  - hs-crp
  - bnp
  - troponin
  - acc-aha-2017
  - esh-2018
  - abpm-task-force
  - carey-resistant-htn
  - aha-acc-pce-2013
  - longitudinal-ehr
  - ehr-synthetic
  - clinical-trial-simulation
pretty_name: HCCAR003 — Synthetic Hypertension & Cardiovascular Risk Dataset (Sample)
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files: hccar003_dataset.parquet

HCCAR003 — Synthetic Hypertension & Cardiovascular Risk Dataset (Sample Preview)

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

A longitudinal hypertension cohort dataset with quarterly visit-level records spanning 7 clinical modules: BP monitoring (office, home, ABPM 24hr/day/night with dipping pattern, central aortic, pulse wave velocity, augmentation index, BP variability), antihypertensive medications (up to 4 drug slots, 8 drug classes — ACEi, ARB, CCB, Thiazide, Beta-blocker, Alpha-blocker, MRA), lifestyle factors (DASH diet, dietary sodium, physical activity METs, sleep, smoking, alcohol, stress), biomarkers (full lipid panel, hs-CRP, creatinine, eGFR with KDIGO CKD stage, UACR with microalbuminuria/macroalbuminuria, electrolytes, glucose, HbA1c, BNP, troponin, ASCVD 10y risk, Framingham), end-organ damage (LVH, LV mass index, LVEF, E/e' ratio, carotid IMT, plaque, retinopathy grade, white matter lesion volume, lacunar infarcts), and MACE outcomes (MI, ischemic / hemorrhagic stroke, TIA, HF hospitalization, AF new onset, CV death, all-cause death, study dropout).

Calibrated benchmark-first against ACC/AHA 2017 Hypertension Guidelines (Whelton et al.), ACC/AHA 2013 Pooled Cohort Equations (Goff et al.), KDIGO 2012 CKD Classification, ABPM Task Force / ESH 2013 Recommendations, AHA/ACC Resistant Hypertension Scientific Statement (Carey et al. 2018), and ESH/ESC 2018 Hypertension Guidelines.

This is the sample preview — 150 patients × 12 quarterly visits over 3 years (1,800 visit records, ~1.1 MB). The full product covers 10,000+ patients × full 10-year follow-up (40 quarterly visits) with extended medication titration histories, multi-cuff measurement protocols, and pre-built scenario configs for SPRINT-style intensive vs standard BP target trials, salt-sensitive hypertension studies, and resistant hypertension subgroup analysis.

Dataset summary

Table Rows (sample) What it contains
hccar003_dataset 1,747 One row per patient × visit. 99 features across 7 clinical modules (demographics carried forward + BP monitoring + medications + lifestyle + biomarkers + end-organ damage + MACE outcomes). 150 unique patients with up to 12 quarterly visits each (some patients drop out due to death or withdrawal)

Provided in both CSV and Parquet. Aggregate to patient-level via groupby('patient_id') for cross-sectional analysis.

Calibration sources

All ten validation metrics target named clinical / regulatory standards:

  • ACC/AHA 2017 Hypertension Guidelines (Whelton et al. 2018) — HTN stage classification (Normal, Elevated, Stage 1, Stage 2, Crisis)
  • ACC/AHA 2013 Pooled Cohort Equations (Goff et al. 2014) — 10-year ASCVD risk (race/sex-specific, baseline column included with known calibration issues — see Limitations)
  • KDIGO 2012 CKD Classification — eGFR-based G1-G5 staging, UACR-based A1-A3 albuminuria categories
  • ABPM Task Force / ESH 2013 (O'Brien et al. 2013) — ambulatory blood pressure monitoring definitions, dipping pattern classification, masked hypertension criteria
  • AHA/ACC Resistant Hypertension Scientific Statement (Carey et al.
    1. — resistant HTN = SBP ≥130/80 on ≥3 antihypertensives
  • ESH/ESC 2018 Hypertension Guidelines (Williams et al. 2018) — European HTN management framework
  • CKD-EPI 2009 (Levey et al. 2009) — eGFR calculation (note: the generator uses the OLD race-coefficient version; the 2021 NKF-ASN refit removed the race coefficient. See Limitations.)
  • Survival analysis monotonicity — CV death ⊆ all-cause death

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 pulse_pressure_equals_sbp_minus_dbp_rate 1.000 0.99 ±0.01 FLOOR Hemodynamic identity
2 map_equals_dbp_plus_pp_third_rate 1.000 0.99 ±0.01 FLOOR Hemodynamic identity
3 ckd_stage_matches_kdigo_egfr_rate 0.999 0.99 ±0.01 FLOOR KDIGO 2012
4 albuminuria_flags_match_uacr_bands_rate 1.000 0.99 ±0.01 FLOOR KDIGO ACR thresholds
5 abpm_dipping_pattern_matches_dip_pct_rate 0.991 0.99 ±0.01 FLOOR ABPM Task Force / ESH 2013
6 resistant_htn_requires_3_drugs_and_sbp_130_rate 1.000 0.99 ±0.01 FLOOR Carey et al. 2018
7 masked_htn_definition_match_rate 1.000 0.99 ±0.01 FLOOR ABPM Task Force
8 cv_death_implies_all_cause_death_rate 1.000 0.99 ±0.01 FLOOR Survival monotonicity
9 mace_event_flag_matches_event_type_rate 1.000 0.99 ±0.01 FLOOR MACE composite endpoint
10 bp_in_physiologic_bounds_rate 1.000 0.99 ±0.01 FLOOR Physiologic plausibility

Schema highlights (99 cols)

Identity & demographics (8 cols, carried per visit)

patient_id (HC-CAR-XXXXXX), visit_number (1-12), visit_date, years_from_baseline, age_at_visit, sex (Male / Female), race_ethnicity (NonHispanic_White / NonHispanic_Black / Hispanic / Asian / Other), htn_stage_baseline (Normal / Elevated / Stage1_HTN / Stage2_HTN / Crisis).

BP monitoring (19 cols)

sbp_office_mmhg, dbp_office_mmhg, pp_office_mmhg, map_office_mmhg, sbp_home_avg_mmhg, dbp_home_avg_mmhg, abpm_sbp_24hr_mmhg, abpm_dbp_24hr_mmhg, abpm_sbp_daytime_mmhg, abpm_sbp_nighttime_mmhg, abpm_dipping_pct, abpm_dipping_pattern (Reverse_dipper / Non-dipper / Dipper / Extreme_dipper), central_aortic_sbp_mmhg, pulse_wave_velocity_ms, augmentation_index_pct, bp_variability_sd_sbp, white_coat_effect_mmhg, white_coat_flag, masked_hypertension_flag.

Medications (12 cols)

drug_class_{1-4}, drug_name_{1-3}, drug_dose_1, n_antihypertensive_drugs (0-4), medication_adherence_pct, bp_response_sbp_mmhg, resistant_htn_flag, side_effect_code (Dry_Cough / Peripheral_Edema / Hypokalemia / Bradycardia / Orthostatic_Hypotension / Hyperkalemia / None), pill_burden_score.

Lifestyle (8 cols)

bmi_kg_m2, dietary_sodium_mg_day, dash_diet_score, physical_activity_mets_hr_wk, sleep_hours_night, smoking_status (Never / Former / Current), alcohol_drinks_week, stress_score.

Biomarkers (22 cols)

total_cholesterol_mg_dl, ldl_cholesterol_mg_dl, hdl_cholesterol_mg_dl, triglycerides_mg_dl, non_hdl_cholesterol_mg_dl, hs_crp_mg_l, creatinine_mg_dl, egfr_ml_min_1_73m2, uacr_mg_g, ckd_stage (G1 / G2 / G3a / G3b / G4 / G5), bun_mg_dl, potassium_meq_l, sodium_meq_l, glucose_fasting_mg_dl, hba1c_pct, uric_acid_mg_dl, bnp_pg_ml, troponin_i_ng_l, ascvd_10yr_risk_pct, framingham_risk_score_pct, microalbuminuria_flag, macroalbuminuria_flag.

End-organ damage (9 cols)

lvh_flag, lv_mass_index_g_m2, lvef_pct, e_e_prime_ratio, carotid_imt_mm, carotid_plaque_flag, retinopathy_grade (0-4), wml_volume_ml, lacunar_infarct_flag.

MACE outcomes (11 cols)

mace_event_flag, mace_event_type (MI / Stroke_Ischemic / Stroke_Hemorrhagic / TIA / HF_Hospitalization / AF_New_Onset / CV_Death / Non_CV_Death / None), mi_flag, stroke_flag, stroke_type, hf_hospitalization_flag, af_new_onset_flag, cardiovascular_death_flag, all_cause_death_flag, study_dropout_flag, dropout_reason (Death / Withdrawal / None).

Comorbidities (7 cols, carried forward)

diabetes_flag, dyslipidemia_flag, osa_flag, statin_use_flag, aspirin_use_flag, family_history_htn_flag, charlson_comorbidity_index.

Suggested use cases

  • BP control prediction — train classifiers/regressors on sbp_office_mmhg and resistant_htn_flag from baseline + medication
    • lifestyle features
  • ABPM interpretation ML — predict dipping pattern, masked HTN from office BP, home BP, and patient features; useful for ABPM-replacement algorithms
  • Medication response prediction — model bp_response_sbp_mmhg given drug class combinations and patient characteristics (uplift modeling for personalized antihypertensive selection)
  • Resistant HTN cohort identification — classifier for resistant_htn_flag for utilization analytics
  • Antihypertensive adherence ML — predict medication_adherence_pct from pill burden, side effects, SES, age
  • ASCVD risk recalibration — train improved 10-year ASCVD models to compare against ACC/AHA PCE (note: the included ascvd_10yr_risk_pct has known calibration issues — useful for derivation studies that explicitly correct PCE bugs)
  • MACE survival ML — Cox / DeepSurv / random survival forests on the MACE outcomes (aggregate per-visit flags to patient level for TTE analysis)
  • End-organ damage progression — model lvh_flag, carotid_imt_mm, wml_volume_ml trajectories given longitudinal BP control
  • CKD progression in HTN — model egfr_ml_min_1_73m2 decline trajectories; useful for predicting CKD-G3 → G4 transitions
  • White coat / masked HTN detection — classifier for office vs ambulatory discrepancy; useful for diagnostic workflow automation
  • Salt sensitivity studies — use dietary_sodium_mg_day and individual BP response to identify salt-sensitive phenotypes
  • Dipping pattern ML — predict nocturnal dipping pattern (Reverse / Non / Normal / Extreme) from office BP, age, OSA, comorbidities; useful for ABPM-free phenotyping
  • Treatment intensification timing ML — predict when next drug should be added from BP trajectory + current regimen
  • Quality improvement / HEDIS analytics — BP control rate measurement, medication intensification audit (HEDIS Controlling High Blood Pressure metric)

Loading examples 

from datasets import load_dataset

ds = load_dataset("xpertsystems/hccar003-sample", split="train")
print(ds.shape)

import pandas as pd
from huggingface_hub import hf_hub_download

df = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar003-sample", "hccar003_dataset.parquet",
    repo_type="dataset",
))

# HTN stage distribution
print(df.groupby("patient_id")["htn_stage_baseline"].first()
      .value_counts(normalize=True).round(3))

# BP trajectory by HTN stage
import pandas as pd
from huggingface_hub import hf_hub_download

df = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar003-sample", "hccar003_dataset.parquet",
    repo_type="dataset",
))

trajectory = (
    df.groupby(["htn_stage_baseline", "visit_number"])["sbp_office_mmhg"]
    .mean().unstack(level=0).round(1)
)
print(trajectory.head(12))

# Resistant HTN identification + drug regimens
import pandas as pd
from huggingface_hub import hf_hub_download

df = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar003-sample", "hccar003_dataset.parquet",
    repo_type="dataset",
))

# Patients with resistant HTN at ANY visit
resistant_pts = df.loc[df["resistant_htn_flag"] == 1, "patient_id"].unique()
print(f"Resistant HTN patients: {len(resistant_pts)} / {df['patient_id'].nunique()}")

# Their typical drug regimens (drug_class_1 distribution)
resist_df = df[df["patient_id"].isin(resistant_pts)]
print(resist_df["drug_class_1"].value_counts().head())

# Aggregate MACE outcomes to patient level
import pandas as pd
from huggingface_hub import hf_hub_download

df = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar003-sample", "hccar003_dataset.parquet",
    repo_type="dataset",
))

# Per-patient MACE summary (per-visit flags need aggregation)
patient_mace = df.groupby("patient_id").agg(
    any_mi=("mi_flag", "max"),
    any_stroke=("stroke_flag", "max"),
    any_hf=("hf_hospitalization_flag", "max"),
    any_af=("af_new_onset_flag", "max"),
    cv_death=("cardiovascular_death_flag", "max"),
    all_cause_death=("all_cause_death_flag", "max"),
    follow_up_yrs=("years_from_baseline", "max"),
).round(3)

print("Patient-level event rates:")
print(patient_mace.mean().round(3))

# ABPM dipping pattern distribution
import pandas as pd
from huggingface_hub import hf_hub_download

df = pd.read_parquet(hf_hub_download(
    "xpertsystems/hccar003-sample", "hccar003_dataset.parquet",
    repo_type="dataset",
))

print("Dipping pattern distribution by HTN stage:")
print(pd.crosstab(df["htn_stage_baseline"], df["abpm_dipping_pattern"],
                  normalize="index").round(3))

Limitations and honest disclosures

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

  • The included Pooled Cohort Equation (ascvd_10yr_risk_pct) is BUGGY. The generator's PCE implementation (lines 122-153) has multiple errors: (a) White Male branch uses 1.764 for both bp_treated and untreated, removing treatment effect; (b) White Female has a misplaced -29.799*1 term; (c) Black Male is missing the published intercept; (d) Black Female has implausible coefficients on ln_sbp (29.2907 vs published 0.295). Result: ASCVD risk values saturate at clip ceiling (75%) for ~93% of patients and at clip floor (1%) for Black Males. Use the column only as a relative-ordering signal, not for absolute 10-year ASCVD estimation. For accurate ASCVD risk, recompute from age, sex, race, total cholesterol, HDL, SBP, BP treatment flag, diabetes, smoking using the published formula (Goff et al. 2014).
  • eGFR uses the OLD 2009 CKD-EPI formula with race coefficient (line 216-217: egfr *= 1.159 if Black). The 2021 NKF-ASN refit REMOVED the race coefficient. If you need the modern formula, recompute from creatinine_mg_dl, age_at_visit, sex without the race multiplier.
  • carotid_plaque_flag formula bug: int(RNG.random() < (0.05 + 0.01*age - 0.5)) effectively gives ZERO carotid plaque to patients under age 45 (the probability term goes negative). For full plaque modeling, use the full product or augment with separate carotid imaging modules.
  • MACE per-visit flags fire on only ONE visit per patient (the visit within 0.13 years of event_time). They are NOT cumulative — the flag does not carry forward after the event. For patient-level MACE prediction, aggregate via groupby('patient_id').max() on the flag columns (see "Loading examples" above). Time-to-event must be derived from the visit_number × 0.25 (quarterly) offset.
  • af_new_onset_flag can fire from MACE event AND from a separate stochastic check (line 480): int(mace_type == 'AF_New_Onset' or (yrs > 5 and sbp_off > 160 and random < 0.005)). So AF can occur outside of the formal MACE event window — by design, reflecting that AF is sometimes diagnosed incidentally.
  • _dropout_at reference in main loop (line 646) checks a dictionary key that's never set by generate_patient_baseline. The branch is dead code; dropout actually fires via study_dropout_flag only. Cosmetic side-effect.
  • Mean SBP ~144 mmHg, mean DBP ~97 mmHg in this sample — higher than typical real-world HTN cohorts (135/85) because the HTN stage distribution skews toward Stage 1/2 (55% of patients). The generator's stage probabilities [Normal: 15%, Elevated: 20%, Stage1: 30%, Stage2: 25%, Crisis: 10%] produce a hypertension- enriched cohort by design (suitable for HTN clinical trials, not general population epidemiology).
  • Race/ethnicity SBP offsets (line 106-109): Black patients have +6 mmHg SBP offset, Asian -2 mmHg. These match published trial observations (e.g., AASK, ALLHAT) but are NOT a complete model of hypertension disparities — they encode only the magnitude offset, not the underlying mechanisms (RAAS responsiveness, salt sensitivity, vascular dysfunction).
  • Visit dropout is independent of clinical state (line 485: dropout_flag = int(dead_v or (random < 0.003))). Real HTN cohort dropout correlates with poor BP control, adverse drug effects, and SES. Treat the sample as informatively-censored data only if you augment with realistic dropout mechanisms.
  • Comorbidities are independent draws (lines 188-194: dm, dys, ckd, osa) — no realistic co-occurrence beyond per-flag base rates. Real cardiometabolic clustering (diabetes + dyslipidemia + obesity + CKD) is much tighter than the generator produces.
  • scipy.stats is imported but unused in active generator code. No external compute dependencies beyond numpy + pandas + tqdm. The scipy distributions (norm, beta, lognorm, weibull_min) appear in the import block but never get called.
  • Masked HTN observed at 0.3-0.5% in sample — much lower than the 10-15% prevalence reported in clinical literature. Generator's wc_effect ~ N(8, 6) and white_coat→office subtraction produces predominantly white-coat phenotype (office > home) rather than masked (home > office). For masked HTN ML research, augment with inverted white-coat scenarios from the full product.
  • Visit count varies by patient — some patients have 12 visits, some have fewer due to dropout. Use groupby('patient_id').size() to check follow-up duration per patient. Treat as unbalanced panel data.
  • Drug-drug interactions and titration are simplified. The drug regimen is fixed at baseline (4 slots, randomly chosen from 7 classes); no realistic titration logic, no switching due to side effects, no addition due to inadequate BP response. For pharmacotherapy intensification ML, use the full product.

The full HCCAR003 product addresses these by corrected ACC/AHA PCE implementation, full 2021 CKD-EPI refit, complete carotid plaque modeling, MACE flag carry-forward for survival analysis, realistic medication titration trajectories, dependent comorbidity sampling, and pre-built scenario configs (SPRINT-style intensive vs standard, salt-sensitive HTN, resistant HTN subgroup). Contact us for the licensed commercial release.

Companion datasets

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

  • HCCAR001 Heart Failure Dataset — chronic HF longitudinal records with GDMT, device therapy, hospitalization, 12 quarterly visits
  • HCCAR002 Acute Myocardial Infarction Dataset — STEMI/NSTEMI/UA with serial troponin kinetics, intervention timing, in-hospital outcomes
  • HCCAR003 Hypertension Dataset (you are here) — longitudinal HTN cohort with ABPM, GDMT, MACE outcomes
  • Healthcare / Neurology (10 SKUs)
  • Insurance & Risk (10 SKUs)
  • Energy & Climate (8 SKUs)
  • Manufacturing (10 SKUs)
  • Oil & Gas (17 SKUs)

Cardiology pairing: HCCAR001 + HCCAR002 + HCCAR003 covers the full HTN→AMI→HF clinical trajectory. Hypertension is the leading modifiable risk factor for AMI (HCCAR002) and HFpEF (HCCAR001 phenotype).

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

Citation 

@dataset{xpertsystems_hccar003_sample_2026,
  author       = {XpertSystems.ai},
  title        = {HCCAR003 Synthetic Hypertension \& Cardiovascular Risk Dataset (Sample Preview)},
  year         = 2026,
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/xpertsystems/hccar003-sample}
}

Contact

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 (HTN stage, resistant HTN, ABPM dipping pattern, KDIGO CKD stages) are sourced from published guidelines; users are responsible for verifying against current ACC/AHA/ESC/KDIGO guidelines for clinical applications. The included Pooled Cohort Equation implementation has known calibration issues — see Limitations.