hccar010-sample / README.md
pradeep-xpert's picture
Upload folder using huggingface_hub
dd8dc08 verified
---
license: cc-by-nc-4.0
language:
- en
tags:
- synthetic-data
- healthcare
- cardiology
- neurology
- stroke
- stroke-prevention
- atrial-fibrillation
- cardiovascular-risk
- framingham
- chads2vasc
- toast-classification
- xpertsystems
pretty_name: "HC-CAR-010 — Stroke Risk Prediction Synthetic Cohort (sample)"
size_categories:
- n<1K
task_categories:
- tabular-classification
- tabular-regression
- time-series-forecasting
---
# HC-CAR-010 — Stroke Risk Prediction Synthetic Cohort
**Sample dataset (500 patients × 144 columns) from the XpertSystems.ai Synthetic Data Factory — Cardiology vertical**
A fully synthetic cohort designed for **stroke risk prediction modeling**,
spanning the complete primary/secondary prevention pipeline: demographics
and family history, comprehensive cardiovascular risk factors, cardiac and
inflammatory/coagulation biomarkers, atrial fibrillation phenotyping with
CHA₂DS₂-VASc/HAS-BLED, carotid + cardiac + neuroimaging, validated risk
scores (Framingham 10-year stroke, SCORE2, Charlson, RRE-90), and acute
stroke outcomes (NIHSS, mRS, Barthel, TOAST classification, thrombolysis,
thrombectomy, recurrence, 1-year mortality).
Built to be **drop-in usable for analytics, modeling, demos, and education**
while remaining 100% synthetic — no real patient data, no PHI, no
re-identification risk.
---
## At a glance
| | |
|---|---|
| **SKU** | HC-CAR-010 |
| **Vertical** | Healthcare → Cardiology/Neurology |
| **Sample size** | 500 patients × 144 columns |
| **Modules** | 10 (Demographics, CV Risk Factors, Biomarkers, Coagulation, AFib, Vascular Imaging, Cardiac Imaging, Neuroimaging, Risk Scores, Outcomes) |
| **Reporting standard** | Framingham/SCORE2/CHA₂DS₂-VASc/TOAST compatible |
| **Format** | CSV |
| **License (sample)** | CC-BY-NC-4.0 |
| **License (full product)** | Commercial — contact XpertSystems.ai |
| **Validation** | **Grade A+ (10.0/10) across all 6 canonical seeds {42, 7, 123, 2024, 99, 1}** |
---
## What makes this dataset useful
Stroke prevention modeling sits at the intersection of cardiology, neurology,
hematology, and behavioral science — and the data spans EHR, registry, lab,
and imaging silos. This synthetic cohort gives you the **full stroke risk
phenome in one tidy table** — including the linked CHA₂DS₂-VASc/HAS-BLED
scoring for the AFib subset, the validated Framingham/SCORE2 risk scores,
TOAST etiology of ischemic strokes, NIHSS/mRS/Barthel post-stroke outcomes,
and structural identities (anticoagulation upgrade for high-risk AFib;
thrombolysis only for ischemic; TOAST only for ischemic) — so you can
prototype models, build training labs, demo dashboards, or teach stroke
epidemiology without paperwork.
Coverage:
- **CV risk factors** — BP (with bp_category staging), lipids (with statin
treatment effect modeling), diabetes, lifestyle, comorbidities (CKD, OSA,
depression, migraine ± aura)
- **Cardiac biomarkers** — NT-proBNP, BNP, troponin I & T, hs-CRP, D-dimer,
fibrinogen, homocysteine, Lp(a), apoB/apoA-1, myeloperoxidase, galectin-3,
ST2 — calibrated to comorbidity (HF, prior MI, diabetes, smoking)
- **Coagulation panel** — INR (warfarin-aware), PTT, platelets, MPV,
antiphospholipid, Factor V Leiden, Protein C/S, antithrombin III
- **AFib phenotyping** — type (paroxysmal/persistent/permanent/long-standing),
detection method, duration, CHA₂DS₂-VASc, CHADS₂, HAS-BLED, LA dimensions,
LAA thrombus, cardioversion, ablation
- **Carotid + vascular imaging** — IMT, plaque, % stenosis (categorical),
vertebral stenosis, ABI, aortic arch plaque grade, CAC score
- **Cardiac imaging** — LVEF, HFrEF/mrEF/pEF classification, NYHA, LVH,
LV mass, E/A, e', PFO, valvular disease
- **Neuroimaging** — Fazekas leukoaraiosis, cerebral microbleed count,
silent lacunar infarcts, brain atrophy GCA, MCA & basilar stenosis
- **Validated risk scores** — Framingham 10-yr stroke, SCORE2 10-yr CVD,
Charlson, RRE-90, derived risk stratum (Low/Moderate/High/Very_High)
- **Acute stroke outcomes** — event flag, ischemic/hemorrhagic/TIA type,
TOAST etiology (Cardioembolic/Large_Artery/Small_Vessel/Other/Undetermined),
time-to-event, NIHSS, mRS 90d, Barthel 90d, DWI lesion volume & location,
thrombolysis, thrombectomy, ICU, LOS, 90-day recurrence, 1-yr mortality
---
## Calibration anchors (industry-grade)
This cohort is calibrated against named registries, guidelines, and trials —
not invented distributions. Selection from the 30-metric scorecard:
| Metric | Sample value (seed 42) | Target range | Source |
|---|---:|---|---|
| Age median | 63.1 yr | 55–70 | Framingham, ARIC |
| Hypertension % | 63.4% | 50–75 | AHA cohorts (enriched) |
| Diabetes % | 27.2% | 18–35 | NHANES ≥40 (enriched) |
| Current smoker % | 28.8% | 18–35 | Stroke risk cohorts |
| CKD any stage % | 45.2% | 35–60 | Lin 2018, ARIC |
| AFib age 70–80 | 14.3% | 10–22 | Rotterdam Study (HF-enriched) |
| AFib overall | 11.4% | 8–18 | High-risk CV cohort |
| CHA₂DS₂-VASc median (AFib) | 4.0 | 2–5 | Stroke-risk AFib cohort |
| Ischemic % of strokes | 86.6% | ≥75% (floor) | GBD 2022: 87% |
| Hemorrhagic % of strokes | 8.5% | 2–18 | GBD 2022: ~10% |
| NIHSS median (ischemic) | 6.0 | 3–11 | SITS-MOST: median 8 acute |
| 1-yr mortality (stroke) | 6.1% | 2–25 | BENCHMARK: 8% post-stroke |
| 90-day recurrence | 1.2% | 1–9 | AHA/ASA ~4% |
| Carotid IMT mean | 0.97 mm | 0.80–1.10 | MESA enriched |
| Carotid plaque % | 49.6% | 30–60 | MESA age >50 |
| Statin % | 56.8% | 40–70 | Secondary prevention enriched |
| Antiplatelet any % | 65.8% | 55–80 | AHA prevention |
| Antihypertensive in HTN | 66.6% | ≥60% (floor) | AHA/ACC 2017 |
| BP control in treated | 55.0% | 45–65 | NHANES |
| Anticoag in AFib | 96.5% | ≥75% (floor) | AHA/ACC/HRS 2023 |
| DOAC share of AFib anticoag | 69.1% | ≥50% (floor) | 2023 AHA/ACC DOAC-preferred |
| LDL mean | 100.3 mg/dL | 85–115 | Statin-treated cohort |
| HDL mean | 47.5 mg/dL | 42–55 | NHANES |
| PFO % | 23.8% | 15–35 | Hagen 1984 autopsy |
| HF % | 23.2% | 12–32 | CV risk cohort enriched |
| Framingham 10-yr risk mean | 12.6% | 8–18 | High-risk cohort |
| TOAST cardioembolic % | 18.3% | 8–30 | Adams 1993 / Petty 1999 |
Full 30-metric scorecard ships in `validation_report.json` and `validation_report.md`.
---
## Files in this sample
```
hccar010_sample/
├── hccar010_sample.csv # 500 patients × 144 columns
├── validation_report.json # full scorecard (machine-readable)
├── validation_report.md # full scorecard (human-readable)
├── sweep_summary.json # 6-seed canonical sweep results
└── README.md # this file
```
---
## Schema (144 columns across 10 modules)
### Module 1 — Demographics (14 cols)
`patient_id`, `age_years`, `sex`, `race_ethnicity`, `education_years`,
`insurance_type`, `urban_rural`, `family_history_stroke_flag`,
`family_history_cad_flag`, `prior_stroke_flag`, `prior_tia_flag`,
`prior_mi_flag`, `prior_cabg_flag`, `prior_pci_flag`
### Module 2 — CV Risk Factors (32 cols)
BP (`systolic_bp_mmhg`, `diastolic_bp_mmhg`, `pulse_pressure_mmhg`,
`bp_category`, `hypertension_flag`, `antihypertensive_flag`,
`antihypertensive_class`, `bp_control_achieved_flag`),
lipids (`total_cholesterol_mg_dl`, `ldl_cholesterol_mg_dl`, `ldl_untreated_mg_dl`,
`hdl_cholesterol_mg_dl`, `triglycerides_mg_dl`, `non_hdl_cholesterol_mg_dl`,
`statin_therapy_flag`, `statin_intensity`), diabetes (`diabetes_type`,
`diabetes_flag`, `hba1c_pct`, `fasting_glucose_mg_dl`,
`diabetes_duration_years`), lifestyle (`smoking_status`, `pack_years`,
`alcohol_drinks_per_week`, `heavy_drinker_flag`,
`physical_activity_mets_week`, `bmi_kg_m2`, `obesity_class`),
comorbidities (`sleep_apnea_flag`, `ckd_stage`, `egfr_ml_min_173m2`,
`depression_flag`, `migraine_flag`, `migraine_with_aura_flag`)
### Module 3 — Cardiac Biomarkers (16 cols)
`bnp_pg_ml`, `nt_probnp_pg_ml`, `troponin_i_ng_ml`, `troponin_t_ng_ml`,
`hs_crp_mg_l`, `d_dimer_ng_ml`, `fibrinogen_mg_dl`, `homocysteine_umol_l`,
`lp_a_mg_dl`, `apob_mg_dl`, `apoa1_mg_dl`, `myeloperoxidase_pmol_l`,
`galectin_3_ng_ml`, `st2_ng_ml`, `creatinine_mg_dl`,
`urine_albumin_creatinine_mg_g`
### Module 4 — Coagulation & Hematology (13 cols)
`inr`, `ptt_seconds`, `platelet_count_k_ul`, `mean_platelet_volume_fl`,
`hemoglobin_g_dl`, `hematocrit_pct`, `antiphospholipid_antibody_flag`,
`factor_v_leiden_flag`, `protein_c_activity_pct`, `protein_s_activity_pct`,
`antithrombin_iii_pct`, `anticoagulant_therapy`, `antiplatelet_therapy`
### Module 5 — Atrial Fibrillation (12 cols)
`afib_flag`, `afib_type` (Paroxysmal/Persistent/Permanent/Long_Standing),
`afib_duration_years`, `afib_detection_method`, `chads2_score`,
`chads2vasc_score`, `hasbled_score`, `left_atrial_diameter_mm`,
`left_atrial_volume_ml`, `left_atrial_appendage_thrombus_flag`,
`cardioversion_flag`, `ablation_flag`
### Module 6 — Carotid & Vascular Imaging (12 cols)
`carotid_imt_right_mm`, `carotid_imt_left_mm`, `carotid_plaque_flag`,
`carotid_stenosis_right_pct`, `carotid_stenosis_left_pct`,
`carotid_stenosis_category` (None/Mild/Moderate/Severe/Occluded),
`vertebral_artery_stenosis_flag`, `ankle_brachial_index`, `pad_flag`,
`aortic_arch_plaque_grade`, `coronary_artery_calcium_score`, `cac_category`
### Module 7 — Cardiac Imaging (11 cols)
`echo_lvef_pct`, `heart_failure_flag`, `heart_failure_type` (HFrEF/HFmrEF/HFpEF),
`nyha_class`, `echo_lv_hypertrophy_flag`, `echo_lv_mass_index_g_m2`,
`echo_e_a_ratio`, `echo_e_prime_cm_s`, `patent_foramen_ovale_flag`,
`valvular_disease_flag`, `valvular_disease_type`
### Module 8 — Neuroimaging (9 cols)
`leukoaraiosis_grade_fazekas`, `cerebral_microbleed_count`,
`silent_lacunar_infarct_count`, `brain_atrophy_grade_gca`,
`mca_stenosis_pct`, `basilar_artery_stenosis_pct`,
`ct_brain_performed_flag`, `mri_brain_performed_flag`, `cta_performed_flag`
### Module 9 — Risk Scores (5 cols)
`framingham_stroke_10yr_risk_pct`, `score2_10yr_cvd_risk_pct`,
`charlson_comorbidity_index`, `rre_90_recurrence_risk_pct`,
`stroke_risk_stratum` (Low/Moderate/High/Very_High)
### Module 10 — Outcomes (18 cols)
`stroke_event_flag`, `stroke_type_at_event` (Ischemic/Hemorrhagic/TIA/NA),
`stroke_etiology_toast` (TOAST 5-class), `time_to_stroke_days`,
`stroke_event_date`, `thrombolysis_tpa_flag`, `thrombectomy_flag`,
`icu_admission_flag`, `hospital_los_days`, `nihss_score`, `mrs_90_day`,
`barthel_index_90_day`, `dwi_lesion_volume_ml`, `dwi_lesion_location`,
`recurrent_stroke_90d_flag`, `mortality_1yr_flag`, `cause_of_death`,
`index_date`
---
## Use cases
1. **Primary stroke prevention modeling** — train classifiers using
demographics, risk factors, biomarkers, imaging → `stroke_event_flag`.
2. **CHA₂DS₂-VASc validation analytics** — verify the implementation,
stratify AFib patients by score, model anticoagulation decisions.
3. **TOAST classification ML** — multi-class etiology models using
AFib + carotid stenosis + lacunar count → ischemic stroke subtype.
4. **Survival analysis** — Cox PH on `time_to_stroke_days` with full
competing-risks setup (recurrence, mortality).
5. **NIHSS → mRS outcome prediction** — acute severity to 90-day
functional outcome (Barthel/mRS).
6. **Carotid stenosis → stroke pipeline** — model the chain from IMT →
plaque → stenosis category → ischemic stroke risk.
7. **AFib anticoagulation appropriateness** — HAS-BLED vs CHA₂DS₂-VASc
net clinical benefit decision modeling.
8. **GDMT gap analytics** — measure preventive therapy gaps (statin %,
antiplatelet %, BP control %) for cohort-level QI dashboards.
9. **Real-world thrombolysis/thrombectomy uptake** — acute stroke
workflow benchmarking.
10. **Teaching & training** — neurology + cardiology residents,
ML-for-healthcare courses, stroke prevention bootcamps.
---
## Loading examples
### pandas
```python
import pandas as pd
df = pd.read_csv("hccar010_sample.csv")
print(df.shape) # (500, 144)
print(df["stroke_risk_stratum"].value_counts())
```
### Hugging Face `datasets`
```python
from datasets import load_dataset
ds = load_dataset("xpertsystems/hccar010-sample")
df = ds["train"].to_pandas()
```
### Stroke prediction model
```python
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score
features = [
"age_years","hypertension_flag","diabetes_flag","afib_flag",
"smoking_status","systolic_bp_mmhg","ldl_cholesterol_mg_dl",
"hdl_cholesterol_mg_dl","hba1c_pct","bmi_kg_m2",
"carotid_imt_right_mm","carotid_plaque_flag",
"framingham_stroke_10yr_risk_pct","chads2vasc_score",
"nt_probnp_pg_ml","hs_crp_mg_l","d_dimer_ng_ml",
]
X = pd.get_dummies(df[features], columns=["smoking_status"])
y = df["stroke_event_flag"]
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.25, random_state=42)
clf = GradientBoostingClassifier(random_state=42).fit(X_tr, y_tr)
auc = roc_auc_score(y_te, clf.predict_proba(X_te)[:, 1])
print(f"AUC: {auc:.3f}")
```
### TOAST etiology stratification
```python
ischemic = df[df["stroke_type_at_event"] == "Ischemic"]
toast = ischemic["stroke_etiology_toast"].value_counts(normalize=True)
print(toast)
# Adams 1993 expected: Large_Artery 16%, Cardioembolic 29%, Small_Vessel 16%,
# Other_Determined 3%, Undetermined 36%
```
### 90-day mRS outcome by acute severity
```python
import seaborn as sns
stroke = df[df["stroke_event_flag"] == 1]
sns.boxplot(data=stroke, x="mrs_90_day", y="nihss_score")
```
### AFib appropriateness audit
```python
afib = df[df["afib_flag"] == 1]
appropriateness = afib.groupby(
pd.cut(afib["chads2vasc_score"], bins=[-1, 1, 2, 9])
)["anticoagulant_therapy"].apply(lambda s: (s != "None").mean())
print(appropriateness)
```
---
## Honest limitations & generator quirks
This is a **commercial synthetic dataset** — not a research-grade simulation
study. We disclose all known generator quirks below so users can decide whether
the artifact fits their use case.
1. **Module 8 (Neuroimaging) uses legacy `np.random.poisson` global state**
instead of the modular `rng`. Two columns are affected:
`cerebral_microbleed_count` and `silent_lacunar_infarct_count`. Per-row
reproducibility for these two columns is **not guaranteed** even with the
same seed (column means are stable; per-patient counts can vary by
±1–2 across runs). This cascades to ~1.7% of `stroke_etiology_toast`
labels (because TOAST gates on `silent_lacunar_infarct_count > 2`).
**Mitigation:** the sample wrapper calls `np.random.seed(seed)` before
generation, which makes the first call in a process deterministic;
distributions are stable across all canonical seeds. The full
commercial product migrates these draws to the modular RNG.
2. **AFib prevalence is HF-enriched.** The generator's stated Rotterdam
Study benchmarks (3.8% age 60–70, 9.5% age 70–80) are upgraded for
HF-positive patients via `_hf_flag_pre`, producing observed prevalence
of ~10–18%. This is **appropriate for a CV high-risk cohort** but **not
appropriate for population-level epidemiology.**
3. **CAC=0 prevalence is ~2% vs MESA ~50%.** The generator uses lognormal
sampling for `coronary_artery_calcium_score` starting at exp(2.0)-1 ≈ 6
minimum, producing very few true-zero CAC scores. MESA in a similar age
range has 40–55% CAC=0. The CAC distribution is calibrated for an
atherosclerosis-enriched cohort, not screening-population CAC
distributions. If you're modeling primary prevention screening, treat
CAC as relatively rather than absolutely valued.
4. **Stroke events are time-to-event over a ~10-year window but the
dataset is cross-sectional.** Each patient has one row. `time_to_stroke_days`
is the days-from-`index_date` to the stroke event (or -1 for no event).
This is a **survival-style cohort** suitable for Cox PH, not a
longitudinal panel with repeated visits.
5. **`thrombolysis_tpa_flag` is gated on `tte_days < 270`** (~9 months).
This represents acute-window eligibility but is broader than the real
4.5-hour clinical window — for clinical realism, treat
`thrombolysis_tpa_flag` as "this patient was administered tPA at some
point during their acute event", not "received within the 4.5h window".
6. **Thrombolysis rate in ischemic strokes is low (~0–5%).** Real-world
IV-tPA rates are 5–15% in modern stroke registries. The synthetic
generator under-models acute treatment uptake; if you're modeling
acute care pipelines, augment with external thrombolysis rate
assumptions.
7. **`stroke_etiology_toast` rule for Cardioembolic is "any AFib"** —
real TOAST classification requires both a high-risk cardiac source
(e.g., AFib) AND absence of competing etiology. The synthetic rule
over-assigns Cardioembolic to AFib patients with carotid stenosis
<50% even if they may not have had an actual cardioembolic mechanism.
For TOAST classifier training, this remains useful (it's a coherent
rule); for epidemiology, it under-counts mixed-mechanism strokes.
8. **`mortality_1yr_flag` for non-stroke patients is 2% background rate**
— the generator assigns a small mortality rate to the entire cohort,
not just stroke patients. In a real follow-up cohort, you would want
`time_to_death_days` separately tracked.
9. **CHA₂DS₂-VASc and HAS-BLED are zero for non-AFib patients.** This is
correct AFib-only scoring behavior, but **filter on `afib_flag == 1`**
when analyzing these scores or you'll dilute with zeros.
10. **Race/ethnicity is not coupled to outcomes.** Real-world stroke
epidemiology shows substantial racial disparities (Black patients
have ~2× higher stroke incidence and worse outcomes — REGARDS study).
The synthetic cohort is intentionally race-blinded to avoid encoding
real-world disparity bias into trainees' models. If you're studying
disparities, use real REGARDS or Get With The Guidelines data.
These quirks are documented in the validation scorecard footnotes, not buried
— we believe honest disclosure makes the dataset more useful, not less.
---
## What you get in the full commercial product
| | Sample (this dataset) | Full product |
|---|---|---|
| Patients | 500 | 50,000+ (configurable) |
| Module 8 RNG | Legacy `np.random` (disclosed) | Migrated to modular `rng` |
| CAC distribution | Atherosclerosis-enriched | Configurable (screening vs enriched) |
| Cohort type | Cross-sectional survival | Optional longitudinal panel |
| Thrombolysis modeling | ~5% (low) | Configurable, modern era (10–15%) |
| TOAST rule sophistication | Single-rule cascade | Multi-etiology probabilistic |
| Validation report | Yes (30 metrics) | Yes + custom scorecard |
| Format | CSV | CSV, Parquet, JSON |
| License | CC-BY-NC-4.0 (non-commercial) | Commercial use license |
| Race-outcome coupling | None (race-blinded) | Configurable disparity profiles |
| Schema export | — | GWTG-Stroke / REGARDS / SITS mapping |
| Support | Community | Email / SLA |
---
## Citation
```bibtex
@dataset{xpertsystems_hccar010_2026,
title = {HC-CAR-010: Stroke Risk Prediction Synthetic Cohort},
author = {{XpertSystems.ai}},
year = {2026},
version= {1.0.0},
url = {https://huggingface.co/datasets/xpertsystems/hccar010-sample},
license= {CC-BY-NC-4.0 (sample); Commercial (full product)},
note = {Calibrated against Framingham Heart Study, GBD 2022, MESA, IST-3, SITS-MOST Registry, Rotterdam Study, NHANES 2019-2020, AHA/ASA 2021 Stroke Prevention Guidelines, AHA/ACC/HRS 2023 AFib Guidelines, TOAST classification (Adams 1993).}
}
```
---
## Contact
- **Email:** [pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai)
- **Web:** [https://xpertsystems.ai](https://xpertsystems.ai)
- **Vertical:** Healthcare / Cardiology
- **SKU catalog:** 10 SKUs shipped in Cardiology (vertical complete), ~75 SKUs across 8 verticals
XpertSystems.ai — synthetic data, calibrated to real-world registries.