Hugging Face's logo

Datasets:
xpertsystems
/
hconc002-sample

Tasks:
Tabular Classification
Tabular Regression
Time Series Forecasting
Languages:
English
Size:
n<1K
Tags:
synthetic-data
healthcare
oncology
lung-cancer
nsclc
sclc
License:
Dataset card Files
xet
 Community
hconc002-sample / README.md
pradeep-xpert's picture
pradeep-xpert
Upload folder using huggingface_hub
eba1d02 verified
preview code
|
raw
history blame contribute delete
21.4 kB
metadata 
license: cc-by-nc-4.0
language:
  - en
tags:
  - synthetic-data
  - healthcare
  - oncology
  - lung-cancer
  - nsclc
  - sclc
  - egfr
  - alk
  - immunotherapy
  - tcga-luad
  - tcga-lusc
  - keynote
  - flaura
  - xpertsystems
pretty_name: HC-ONC-002 — Lung Cancer Synthetic Cohort (sample)
size_categories:
  - n<1K
task_categories:
  - tabular-classification
  - tabular-regression
  - time-series-forecasting

HC-ONC-002 — Lung Cancer Synthetic Cohort

Sample dataset (500 patients × 116 columns) from the XpertSystems.ai Synthetic Data Factory — Oncology vertical, SKU 2

A fully synthetic, multimodal lung cancer cohort spanning the complete clinical pathway: smoking-stratified histology (NSCLC adeno/squamous/large cell + SCLC limited/extensive), AJCC 8th Edition T/N/M staging with site- specific metastases, comprehensive molecular biomarkers (EGFR with variant subtypes, ALK/ROS1 fusions, KRAS with G12C breakout, BRAF V600E, MET ex14, RET, NTRK, HER2, STK11, KEAP1, TP53), PD-L1 TPS+CPS scoring, TMB, MSI, treatment protocols across the IO/TKI era (surgery+adjuvant, SBRT, CCRT+ durvalumab, targeted TKIs, chemo-IO combinations), RECIST treatment response with pseudoprogression/hyperprogression flags, multimodal imaging (PET SUV/ MTV, ctDNA detection+VAF), IHC markers (TTF-1, p40, synaptophysin), adverse events including irAE phenotyping, and survival outcomes (PFS/OS with Weibull-derived event times).

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-ONC-002
Vertical Healthcare → Oncology (SKU 2)
Sample size 500 patients × 116 columns
Modules 9 (Demographics, Histology+Staging, Molecular, Treatment, Response+Survival, Imaging+Pathology, Comorbidities, Adverse Events, Identifiers)
Standards AJCC 8th Edition, NCCN NSCLC/SCLC 2024, RECIST 1.1, CTCAE v5
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

Lung cancer data lives across SEER (population incidence/survival, no molecular), TCGA LUAD/LUSC (deep genomics but n<1,000), clinical trial datasets (FLAURA/ALEX/KEYNOTE/CheckMate — tightly restricted), and real-world commercial datasets (Flatiron, COTA — expensive). This synthetic cohort gives you the full lung cancer molecular+treatment+ outcomes phenome in one tidy table with realistic dependencies:

  • ✅ Smoking ↔ histology coupling — never-smokers are ~60-70% adeno, current smokers more diverse
  • ✅ Adeno ↔ EGFR/ALK coupling — EGFR mutations 14-26% in adeno vs <2% in squamous; ALK 3-7% in adeno vs <1% elsewhere
  • ✅ EGFR/ALK/ROS1 mutual exclusivity (0 co-occurrences enforced)
  • ✅ Stage IV EGFR+ NSCLC → 100% TKI (NCCN Class I structural identity)
  • ✅ SCLC ↔ TP53 coupling — TP53 mutation ~85-94% in SCLC (matches George 2015)
  • ✅ PD-L1 distribution with realistic spikes at 0%, 1-49%, ≥50%, 100%
  • ✅ OS ≥ PFS always (0 violations across cohort)
  • ✅ Treatment-specific survival calibration — FLAURA EGFR osi PFS ~19 mo, ALEX alectinib PFS ~35 mo, KEYNOTE-024 pembro PFS ~14 mo
  • ✅ irAE only in IO-treated patients (0 violations)
  • ✅ IHC marker fidelity — TTF-1+ only in adeno (75%), p40+ only in squamous (90%), synaptophysin+ only in SCLC (85%)

Coverage spans:

  • NSCLC + SCLC combined with smoking-stratified histology assignment
  • AJCC 8th Edition staging (IA/IB/IIA/IIB/IIIA/IIIB/IIIC/IVA/IVB) with T1a-T4 sub-staging, N0-N3 nodal staging, M0/M1a/M1b/M1c
  • Site-specific metastasis flags — brain, bone, liver, adrenal
  • Comprehensive molecular profile — EGFR (Exon19del/L858R/T790M/Exon20ins/ Other), ALK fusions (EML4/KIF5B/Other), ROS1 fusions, KRAS (G12C/G12V/G12D/ G13C/G12A), BRAF (V600E/non-V600E), MET ex14, RET, NTRK, HER2, STK11, KEAP1, TP53
  • Immunooncology biomarkers — PD-L1 TPS + CPS with categorization, TMB high flag, MSI status
  • Treatment regimens — surgery types (lobectomy/segmentectomy/wedge/ VATS/robotic), SBRT, CCRT, IMRT, chemo (cisplatin-pemetrexed, carbo- paclitaxel, etoposide-platinum), IO (pembrolizumab, atezolizumab, durvalumab, nivolumab+ipilimumab), TKIs (osimertinib, alectinib, brigatinib, lorlatinib, entrectinib, sotorasib, adagrasib, dabrafenib- trametinib, tepotinib, capmatinib, selpercatinib), bevacizumab, adjuvant osimertinib (ADAURA-style)
  • RECIST treatment response — CR/PR/SD/PD with ORR/DCR, time-to-response, duration-of-response, CT response % change
  • Pseudoprogression + hyperprogression flags in IO-treated patients
  • Liquid biopsy — ctDNA detection, VAF%, clearance flag
  • Multimodal imaging — PET SUV-max, MTV
  • IHC panel — TTF-1, p40, synaptophysin/CD56
  • Survival outcomes — PFS/OS with Weibull-derived event times, treatment-specific lambda calibration (FLAURA, ALEX, KEYNOTE-024, PACIFIC, IMpower133)
  • Adverse events — irAE type (pneumonitis, colitis, hepatitis, endocrinopathy, dermatitis) with grade, chemo AEs (nausea, neuropathy, cytopenias), G-CSF use, hospitalization

Calibration anchors (industry-grade)

This cohort is calibrated against named registries, guidelines, and trials — not invented distributions. Selection from the 31-metric scorecard:

Metric Sample value (seed 42) Target range Source
Mean age 66.9 yr 62–72 SEER lung cancer
Female % 48.2% 40–56 SEER ~47%
Never smoker % 15.0% 10–25 SEER ~15-20%
Current smoker % 37.2% 30–50 SEER
Adenocarcinoma % 40.0% 32–48 SEER ~40-45%
Squamous % 25.2% 20–33 SEER ~25-30%
SCLC % 29.6% 20–38 Cohort over-enriched vs SEER 13% (disclosed)
Adeno in never-smokers 57.3% 50–90 SEER ~70-85%
Stage IV in NSCLC 44.6% 35–55 SEER ~40-50%
EGFR in adeno 19.0% 10–30 TCGA LUAD ~15%; LCMC ~17%
ALK in adeno 4.5% 2.5–8 Literature ~5-7%
KRAS in adeno 27.0% 14–32 TCGA LUAD ~30%
KRAS G12C in KRAS+ 27.8% 25–50 CodeBreaK 100
PD-L1 zero % 26.6% 22–38 KEYNOTE-024 ~30%
PD-L1 ≥50% 54.0% 40–60 Enriched cohort
TTF-1+ in adeno 73.5% 60–85 Bishop 2010 ASCP
p40+ in squamous 91.3% ≥80% (floor) Bishop 2012 ASCP
TKI in Stage IV EGFR+ NSCLC 100% ≥90% (floor) NCCN Class I
Surgery in early NSCLC 67.8% 50–75 NCDB
CCRT in locally advanced 57.1% 40–80 PACIFIC era
OS median (overall) 15.85 mo 12–22 Mixed cohort
ORR (overall) 46.6% 35–55 Mixed treatment cohort
ECOG 0-1 % 70.8% 60–80 NCCN-era trials
irAE in IO-treated 27.6% 20–40 CheckMate-227
Brain mets in Stage IV NSCLC 29.9% 22–42 Sorensen 1988, Schouten 2002
Bone mets in Stage IV NSCLC 36.3% 28–48 NSCLC autopsy series
TP53 in SCLC 85.1% 75–100 George 2015
ctDNA detection in advanced NSCLC 81.6% ≥70% (floor) Guardant360

Full 31-metric scorecard ships in validation_report.json and validation_report.md.

Files in this sample 

hconc002_sample/
├── hconc002_sample.csv        # 500 patients × 116 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 (116 columns across 9 modules)

Module 1 — Identifiers & Dates (3 cols)

patient_id, site_id, diagnosis_date

Module 2 — Demographics (15 cols)

age_at_diagnosis, sex, race, insurance, smoking_status, pack_years, cigarettes_per_day, smoking_duration_years, years_since_quitting, low_dose_ct_screening_history, second_hand_smoke_exposure, occupational_exposure, radon_exposure_flag, family_history_lung_cancer, bmi

Module 3 — Histology & Staging (18 cols)

histology_primary (Adenocarcinoma/Squamous_Cell/Large_Cell/SCLC_Limited/ SCLC_Extensive), histology_subtype, clinical_stage (IA/IB/IIA/IIB/IIIA/ IIIB/IIIC/IVA/IVB or truncated SCLC "Limi"/"Exte" — see Limitations #1), t_stage, n_stage, m_stage, tumor_size_cm, tumor_location, tumor_laterality, pleural_invasion_flag, vascular_invasion_flag, lymphovascular_invasion_flag, satellite_nodule_flag, brain_metastasis_flag, bone_metastasis_flag, liver_metastasis_flag, adrenal_metastasis_flag, metastasis_sites

Module 4 — Molecular Biomarkers (18 cols)

egfr_mutation, alk_fusion, ros1_fusion, kras_mutation, braf_mutation, met_exon14_skip, ret_fusion, ntrk_fusion, her2_alteration, stk11_mutation, keap1_mutation, tp53_mutation, pd_l1_tumor_proportion_score, pd_l1_combined_positive_score, pd_l1_category, tmb_mutations_per_mb, tmb_high_flag, microsatellite_status

Module 5 — Treatment (14 cols)

treatment_regimen, targeted_therapy, immunotherapy_agent, chemotherapy_regimen, surgery_type, surgical_margin_status, radiation_type, radiation_dose_gy, treatment_cycles_completed, treatment_adherence_pct, dose_reduction_flag, bevacizumab_flag, adjuvant_chemotherapy_flag, adjuvant_osimertinib_flag

Module 6 — Response & Survival (15 cols)

progression_free_survival_months, pfs_event_flag, overall_survival_months, os_event_flag, time_to_treatment_failure_months, best_overall_response, objective_response_flag, disease_control_flag, time_to_response_months, duration_of_response_months, ct_response_pct_change, pseudoprogression_flag, hyperprogression_flag, next_line_therapy_flag, ldh_at_progression_u_l

Module 7 — Imaging & Pathology (9 cols)

pet_ct_suv_max, pet_ct_mtv_ml, ctdna_detection_flag, ctdna_vaf_pct, ctdna_clearance_flag, pathology_grade, ihc_ttf1, ihc_p40, ihc_synaptophysin_cd56

Module 8 — Comorbidities (16 cols)

ecog_performance_status, fev1_pct_predicted, dlco_pct_predicted, copd_flag, copd_gold_stage, cardiovascular_disease_flag, diabetes_flag, hypertension_flag, prior_malignancy_flag, charlson_comorbidity_index, albumin_g_dl, ldh_baseline_u_l, hemoglobin_g_dl, neutrophil_lymphocyte_ratio, platelet_lymphocyte_ratio, c_reactive_protein_mg_l

Module 9 — Adverse Events (8 cols)

irae_flag, irae_type, irae_grade, nausea_grade, peripheral_neuropathy_grade, cytopenias_grade, hospitalization_flag, g_csf_use_flag

Use cases

  1. Histology classification models — train classifiers using smoking history, demographics, imaging features → adeno/squamous/SCLC subtype.
  2. EGFR/ALK/KRAS biomarker prediction — clinical+demographic features → likelihood of actionable mutation; benchmark precision-medicine referral logic.
  3. Treatment selection modeling — NCCN guideline-concordance scoring (TKI for driver mutations, IO for PD-L1≥50%, CCRT+durvalumab for locally advanced).
  4. Survival prediction — Cox PH on PFS/OS with stage + molecular + treatment covariates; treatment-specific landmark analyses.
  5. RECIST response prediction — multimodal features → ORR / pCR / hyperprogression risk.
  6. PD-L1 distribution analytics — score distribution modeling for trial inclusion criteria.
  7. Liquid biopsy modeling — ctDNA detection probability by stage + tumor burden; VAF dynamics.
  8. Immune-related adverse event prediction — risk stratification by IO agent + clinical features.
  9. Real-world data benchmarking — quasi-experimental analyses with treatment arm comparisons.
  10. Teaching & training — oncology fellows, lung cancer multidisciplinary conferences, ML-for-healthcare courses.

Loading examples

pandas 

import pandas as pd
df = pd.read_csv("hconc002_sample.csv")
print(df.shape)        # (500, 116)
print(df["histology_primary"].value_counts())
print(df.groupby("clinical_stage")["overall_survival_months"].median())

Hugging Face datasets 

from datasets import load_dataset
ds = load_dataset("xpertsystems/hconc002-sample")
df = ds["train"].to_pandas()

Driver mutation classification 

from sklearn.ensemble import GradientBoostingClassifier
from sklearn.model_selection import train_test_split

# EGFR vs no-EGFR in adenocarcinoma
adeno = df[df["histology_primary"] == "Adenocarcinoma"].copy()
adeno["egfr_pos"] = (adeno["egfr_mutation"] != "None").astype(int)
features = ["age_at_diagnosis", "sex", "smoking_status", "pack_years",
            "race", "family_history_lung_cancer", "tumor_size_cm",
            "tp53_mutation", "pd_l1_tumor_proportion_score"]
X = pd.get_dummies(adeno[features])
y = adeno["egfr_pos"]

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)
print(f"AUC features: {sorted(zip(X.columns, clf.feature_importances_), key=lambda x: -x[1])[:5]}")

Survival analysis by treatment regimen 

from lifelines import KaplanMeierFitter
import matplotlib.pyplot as plt

stage_iv = df[df["clinical_stage"].isin(["IVA","IVB"])].copy()
kmf = KaplanMeierFitter()
for reg, sub in stage_iv.groupby("treatment_regimen"):
    if len(sub) < 5: continue
    kmf.fit(sub["overall_survival_months"], event_observed=sub["os_event_flag"], label=reg)
    kmf.plot_survival_function()
plt.title("OS by Treatment Regimen — Stage IV NSCLC")
plt.show()

NCCN guideline-concordance audit 

# NCCN: TKI for EGFR+ Stage IV NSCLC
nsclc_iv_egfr = df[(~df["histology_primary"].isin(["SCLC_Limited","SCLC_Extensive"]))
                   & (df["clinical_stage"].isin(["IVA","IVB"]))
                   & (df["egfr_mutation"] != "None")]
tki_rate = (nsclc_iv_egfr["treatment_regimen"] == "Targeted_TKI").mean()
print(f"TKI in Stage IV EGFR+ NSCLC: {tki_rate:.1%} (NCCN target ≥90%)")

# NCCN: CCRT+durvalumab for locally advanced unresectable NSCLC
locally_adv = df[df["clinical_stage"].isin(["IIIA","IIIB","IIIC"])
                 & (~df["histology_primary"].isin(["SCLC_Limited","SCLC_Extensive"]))]
ccrt_rate = (locally_adv["treatment_regimen"] == "CCRT").mean()
print(f"CCRT in locally advanced NSCLC: {ccrt_rate:.1%}")

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. SCLC stage labels are truncated to 4 characters. Due to a fixed-length string dtype, when the generator assigns clinical_stage = "Limited" or "Extensive" for SCLC patients (after initially populating with NSCLC labels like "IIIC"), the strings are truncated to "Limi" and "Exte". Downstream impact: the m_stage calculation uses np.isin(stage, ["IVA","IVB","Extensive"]) — "Exte" doesn't match "Extensive", so SCLC_Extensive patients incorrectly get assigned m_stage = "M0" and no metastasis sites, despite Extensive SCLC being metastatic by definition. The ctDNA detection rate is also lower in SCLC_Extensive patients (gets ~35% non-advanced rate instead of ~82% advanced rate). The wrapper's metrics use NSCLC-only subsets for metastasis and ctDNA computations to avoid contaminating analyses. Full product fixes the dtype.

  2. SCLC is over-represented at ~30% of cohort vs SEER ~13%. Generator's histology probabilities assign SCLC 27-32% across smoking strata. This is a design choice for a cohort enriched in advanced disease, appropriate for SCLC-focused modeling but not appropriate for population-level epidemiology. For SEER-calibrated SCLC fraction (~13%), sub-sample or re-weight the SCLC subset.

  3. Module 3 (histology assignment) uses np.random.choice (legacy global state) at lines 148, 152, 156 instead of the modular rng. The wrapper mitigates by calling np.random.seed(seed) before generation, but this means per-row histology values are deterministic only for the first call in a process. Distributions are stable across all canonical seeds. Full product migrates these draws to the modular RNG.

  4. CCI calculation has a typo: hypertension contribution multiplied by 0. Line 680 reads htn * 0 instead of htn, effectively excluding hypertension from the Charlson Comorbidity Index sum. Observed CCI mean is ~1.5 (would be ~2.1 with HTN included). The hypertension_flag column is still correctly populated — only the CCI summary metric is affected.

  5. EGFR/ALK/ROS1 are forced mutually exclusive (generator design). This is biologically accurate (true co-occurrence is exceedingly rare) but means compound-driver patients are not represented.

  6. Stage IV EGFR+ NSCLC → 100% TKI assignment is enforced (no chemo-only stage IV EGFR+ patients). NCCN-concordant but real-world ~85-92% receive TKI first-line; the remaining 8-15% receive chemo for reasons like ECOG ≥3, T790M-only mutation, or patient preference — not modeled here.

  7. PD-L1 TPS uses a spike-mixture distribution — spike at 0% (28%), continuous 1-49% (22%), continuous 50-99% (20%), spike at 100% (30%). This produces the characteristic bimodal distribution seen in IO trials but slightly over-represents TPS≥50% (50%) compared to KEYNOTE-024 screening population (30%). Cohort is enriched in IO-eligible patients.

  8. Treatment-specific survival lambdas are point-calibrated to single trials (FLAURA, ALEX, KEYNOTE-024, PACIFIC, IMpower133). Real-world survival distributions show wider variance and include trial-ineligible patients with worse outcomes. Cohort survival skews trial-ish.

  9. Adjuvant osimertinib (ADAURA) flag is independent of EGFR mutation status — the generator assigns adjuvant_osimertinib_flag = 1 with probability 0.80 for early-stage EGFR+ post-surgery patients, but does not block assignment for EGFR-negative patients. Filter on egfr_mutation != "None" before using this flag for ADAURA-style analyses.

  10. Race/ethnicity is not coupled to molecular biomarkers. Real lung cancer epidemiology shows substantial racial differences (EGFR in Asian never-smokers ~50% vs White ~15%; KRAS in White smokers higher than Asian). The synthetic cohort is intentionally race-blinded in molecular assignment to avoid encoding real-world disparity bias into trainees' models. If you're studying disparities, use real LCMC or TCGA-LUAD data.

  11. scipy.stats is NOT imported (clean — no dead imports in this generator), unlike HCONC001.

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 15,000+ (configurable)
SCLC stage truncation "Limi"/"Exte" bug (disclosed) Fixed to "Limited"/"Extensive"
SCLC fraction ~30% (over-enriched) Configurable (SEER 13% → enriched 30%)
Histology RNG Legacy np.random (disclosed) Migrated to modular rng
CCI calculation HTN excluded (bug) Full Charlson
Adjuvant osimertinib gating EGFR-independent Gated on EGFR+
Race-biomarker coupling None (race-blinded) Configurable LCMC-calibrated
Validation report Yes (31 metrics) Yes + custom scorecard
Format CSV CSV, Parquet, JSON
License CC-BY-NC-4.0 (non-commercial) Commercial use license
Schema mapping — SEER / NCDB / TCGA-LUAD-LUSC / Flatiron
Longitudinal extension No Optional treatment-line trajectory
Support Community Email / SLA

Citation 

@dataset{xpertsystems_hconc002_2026,
  title  = {HC-ONC-002: Lung Cancer Synthetic Cohort},
  author = {{XpertSystems.ai}},
  year   = {2026},
  version= {1.0.0},
  url    = {https://huggingface.co/datasets/xpertsystems/hconc002-sample},
  license= {CC-BY-NC-4.0 (sample); Commercial (full product)},
  note   = {Calibrated against SEER lung cancer 2017-2021, TCGA LUAD/LUSC, NCCN NSCLC/SCLC Guidelines 2024, AJCC 8th Edition, FLAURA (Soria 2018), ALEX (Peters 2017), CheckMate-816/9LA (Forde 2022, Paz-Ares 2021), KEYNOTE-024/189/407 (Reck 2016, Gandhi 2018, Paz-Ares 2018), IMpower133 (Horn 2018), PACIFIC (Antonia 2017), CodeBreaK 100 (Skoulidis 2021), Guardant360.}
}

Contact

XpertSystems.ai — synthetic data, calibrated to real-world registries.