xpertsystems/hconc004-sample

HC-ONC-004 — Colorectal Cancer Synthetic Cohort

Sample dataset (500-patient primary cohort + ~3,300-row CEA longitudinal panel) from the XpertSystems.ai Synthetic Data Factory — Oncology vertical, SKU 4

A fully synthetic colorectal cancer cohort spanning the complete clinical pathway: AJCC 8th Edition T/N/M staging across colon + rectum subsites, comprehensive molecular markers (MSI/MMR, KRAS codons 12 & 13, NRAS, BRAF V600E with MSI-H enrichment, HER2 IHC + amplification, PIK3CA, TP53, APC, SMAD4, TMB, PD-L1 CPS, NTRK/RET fusions, ctDNA with VAF), surgical outcomes (R-status, anastomotic leak, lymph node harvest with NCCN adequacy, operative time, EBL, ICU/LOS/readmission, CRM/DRM for rectal, stoma formation), chemo- therapy regimens (MOSAIC/FIRE-3/KEYNOTE-177-era — FOLFOX/CAPOX/FOLFIRI/ FOLFOXIRI, anti-EGFR cetuximab/panitumumab, anti-VEGF bevacizumab, IO with pembrolizumab/nivolumab+ipilimumab, BEACON-CRC for BRAF V600E, larotrectinib for NTRK), RECIST response with depth-of-response, CEA dynamics (baseline, nadir, response/progression flags), survival endpoints (OS/DFS/PFS/recurrence with site), QoL (EORTC QLQ-C30, LARS for rectal), and a variable-length CEA longitudinal panel (18 timepoints over 10 years, truncated by OS).

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-004
Vertical	Healthcare → Oncology (SKU 4)
Sample size	500-patient primary × 105 columns + ~3,300-row CEA panel × 4 cols
Follow-up	Up to 18 CEA timepoints (variable per patient — depends on OS)
Standards	AJCC 8th Edition, NCCN CRC 2024, NCCN Rectal 2024, ESMO 2022
Format	CSV (cohort + longitudinal CEA)
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

CRC data is uniquely fragmented: SEER provides population-level incidence and overall survival but lacks treatment detail and molecular profiles; TCGA COADREAD has deep genomics but n=633; clinical trial datasets (FIRE-3, MOSAIC, KEYNOTE-177, BEACON-CRC) are restricted; real-world commercial datasets (Flatiron, ConcertAI, COTA) are expensive. This synthetic cohort gives you the full CRC phenome in one tidy table with realistic dependencies preserved:

• ✅ AJCC stage ↔ tumor burden coupling — T1-T4 size cascades, N0-N2 positive node ratios
• ✅ MSI-H ↔ stage inversion — MSI-H prevalence ~17% overall but ~3-9% in Stage IV (KEYNOTE-177 calibration)
• ✅ BRAF V600E enriched in MSI-H (~10-19% in MSI-H vs ~3-5% in MSS)
• ✅ KRAS/NRAS/BRAF mutual exclusivity (RAS pathway gating)
• ✅ Treatment selection causally driven — Pembrolizumab only in MSI-H, BEACON-CRC only in BRAF V600E, anti-EGFR only in RAS WT (or MSI-H exception), larotrectinib only in NTRK fusion-positive
• ✅ Adjuvant chemo gated on Stage II-III + R0 (NCCN concordance)
• ✅ Anastomotic leak by surgical approach — open vs laparoscopic vs robotic risk modulation
• ✅ CEA dynamics tied to RECIST response — CR/PR patients show nadir drop to 5-45% of baseline; PD shows rise

Coverage spans:

• AJCC 8th Edition staging (I, IIA, IIB, IIC, IIIA, IIIB, IIIC, IVA, IVB, IVC) with T/N/M sub-staging and pathologic T/N post-surgery
• Anatomic subsites — 11 colon + rectum subsites (Cecum through Lower Rectum); subsite-driven surgical procedure selection
• Stage IV metastasis detail — liver/lung/peritoneal/brain mets flags, liver met count category, liver resectability (resectable/potentially/ unresectable), peritoneal carcinomatosis index (PCI), synchronous vs metachronous mets
• Molecular markers — MSI status (MSI-High/MSI-Low/MSS), MMR (dMMR/pMMR), MLH1 methylation, KRAS codon 12 + 13 with all common variants (G12D/G12V/ G12C/G12A/G12S/G12R/G13D), NRAS (Q61K/Q61R/G12C), BRAF V600E, HER2 IHC + amplification, PIK3CA (Exon9/Exon20), TP53, APC, SMAD4, TMB, PD-L1 CPS, NTRK/RET fusions, ctDNA detection + VAF, CEA baseline
• Surgical outcomes — 12 procedure types (Right_Hemicolectomy through Local_Excision), Open/Laparoscopic/Robotic approach, R0/R1/R2 status, LN harvested + positive + ratio, anastomotic leak (A/B/C grade), wound infection, ileus, laparoscopic→open conversion, operative time, EBL, ICU, LOS, 30d readmission, CRM/DRM for rectal, neoadjuvant flag, pCR flag, perforation, stoma formation + reversal
• Chemotherapy — 20+ regimens calibrated to PFS literature (mFOLFOX6 10.6mo, FOLFOXIRI 12mo, Pembrolizumab 16.5mo, BEACON-CRC 4.3mo, etc.) with adjuvant + palliative gating
• RECIST response — CR/PR/SD/PD with depth-of-response %, CEA nadir, CEA response/progression flags, dose reduction, treatment discontinuation reasons
• Toxicities — oxaliplatin neuropathy grade, febrile neutropenia, hand-foot syndrome grade, bevacizumab hypertension, anti-EGFR skin toxicity
• Survival endpoints — OS, DFS, PFS, recurrence flag + site (liver/lung/ local/peritoneal/nodal/multi), time-to-recurrence, vital status
• QoL — EORTC QLQ-C30, Low Anterior Resection Syndrome (LARS) for rectal
• CEA longitudinal — variable-length panel (3-16 visits per patient) at fixed timepoints (0, 3, 6, 9, 12, 18, 24, 30, 36, 42, 48, 54, 60, 72, 84, 96, 108, 120 months), truncated by OS

---

Calibration anchors (industry-grade)

This cohort is calibrated against named registries, guidelines, and trials — not invented distributions. Selection from the 34-metric scorecard:

Metric	Sample value (seed 42)	Target range	Source
Mean age	67.3 yr	62–72	SEER CRC
Female %	44.4%	40–55	SEER
Lynch syndrome %	2.2%	1.5–5	Hampel 2008
Stage I %	21.8%	15–26	SEER ~20%
Stage IV combined	24.8%	20–30	SEER ~22-25%
Rectum %	26.4%	20–35	SEER ~28%
Liver mets in Stage IV	72.6%	60–82	Engstrand 2018
Synchronous mets in Stage IV	70.2%	50–78	Real-world
MSI-H overall	17.0%	14–24	TCGA COADREAD
MSI-H in Stage IV	3.2%	1.5–12	KEYNOTE-177 ~4-5%
KRAS mutation	41.8%	38–50	TCGA ~43%
KRAS G12C in KRAS+	13.4%	10–25	KRYSTAL-1
RAS WT	52.4%	42–58	TCGA ~50%
BRAF V600E	4.2%	3–10	Literature ~8% (cohort 5-7%)
HER2 amplification	7.2%	3–12	Literature ~5%
PIK3CA	21.4%	16–25	TCGA ~20%
TP53 mutation	56.6%	52–65	TCGA ~60%
APC mutation	83.2%	78–92	TCGA ~80-85%
KRAS/NRAS exclusivity	100%	≥100% (floor)	Structural
LN harvest ≥12	94.2%	≥80% (floor)	NCCN adequacy
R0 resection	85.7%	75–92	NCDB
Anastomotic leak	5.1%	3–9	Modern era
Neoadjuvant in rectal II-III	79.7%	60–90	NCCN
pCR in neoadjuvant	15.7%	10–25	MERCURY
Adjuvant in Stage III	73.6%	55–85	NCDB
Palliative chemo in Stage IV	100%	≥95% (floor)	NCCN
Anti-EGFR in RAS WT	56.0%	35–70	FIRE-3 era
Pembrolizumab in MSI-H only	100%	≥100% (floor)	KEYNOTE-177
ORR (palliative)	39.9%	32–60	Mixed regimen
Stage OS monotonic	100%	≥100% (floor)	Structural

Full 34-metric scorecard ships in validation_report.json and validation_report.md.

---

Files in this sample

hconc004_sample/
├── hconc004_sample.csv               # 500 patients × 105 columns (cohort)
├── hconc004_cea_longitudinal.csv     # ~3,300 rows × 4 columns (CEA panel)
├── 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

The two tables join on patient_id. The CEA longitudinal panel has variable rows per patient (3-16, median ~6) — depends on OS truncation. Columns: patient_id, timepoint_months, cea_ng_ml, assessment_type.

---

Schema (105 columns in cohort + 4 columns in CEA panel)

Cohort: Demographics (12 cols)

patient_id, age_at_diagnosis, sex, race_ethnicity, bmi_kg_m2, smoking_status, diabetes_flag, family_history_crc_flag, lynch_syndrome_flag, lynch_gene (MLH1/MSH2/MSH6/PMS2/None), ecog_performance_status, diagnosis_date

Cohort: Staging (16 cols)

ajcc_stage_group (I/IIA/IIB/IIC/IIIA/IIIB/IIIC/IVA/IVB/IVC), clinical_t_stage, clinical_n_stage, clinical_m_stage, pathologic_t_stage, pathologic_n_stage, tumor_site (Colon/Rectum), tumor_subsite (11 subsites), liver_metastasis_flag, liver_metastasis_count_category, liver_resectability, lung_metastasis_flag, peritoneal_carcinomatosis_flag, peritoneal_carcinomatosis_index, synchronous_metastasis_flag, tumor_deposits_flag

Cohort: Molecular Markers (20 cols)

msi_status, mmr_status, mlh1_promoter_methylation_flag, kras_codon12_mutation, kras_codon13_mutation, nras_mutation, ras_status_combined, braf_v600e_status, pik3ca_mutation, tp53_mutation, apc_mutation, smad4_status, her2_ihc_score, her2_status, tmb_mutations_per_mb, pdl1_combined_positive_score, ctdna_detected_flag, ctdna_vaf_pct, ntrk_fusion_flag, ret_fusion_flag, cea_baseline_ng_ml

Cohort: Surgery (24 cols)

surgery_intent, surgery_procedure, surgical_approach, r_status, lymph_nodes_harvested, lymph_nodes_positive, lymph_node_ratio, anastomotic_leak_flag, anastomotic_leak_grade, wound_infection_flag, ileus_flag, conversion_to_open_flag, operative_time_minutes, estimated_blood_loss_ml, icu_admission_flag, hospital_los_days, readmission_30d_flag, circumferential_resection_margin_positive_flag, distal_resection_margin_mm, neoadjuvant_therapy_flag, pathologic_complete_response_flag, tumor_perforation_flag, stoma_formation_flag, stoma_reversal_flag

Cohort: Chemotherapy (22 cols)

adjuvant_chemo_flag, adjuvant_regimen, adjuvant_cycles_planned, adjuvant_cycles_completed, adjuvant_dose_intensity_pct, palliative_chemo_flag, chemotherapy_regimen_line1, recist_best_response, recist_depth_of_response_pct, cea_nadir_ng_ml, cea_nadir_timing_weeks, cea_response_flag, cea_progression_flag, cycles_completed, dose_reduction_flag, treatment_discontinuation_reason, oxaliplatin_neuropathy_grade, febrile_neutropenia_flag, hand_foot_syndrome_grade, bevacizumab_hypertension_flag, anti_egfr_skin_toxicity_grade, conversion_surgery_flag

Cohort: Survival (10 cols)

overall_survival_months, disease_free_survival_months, progression_free_survival_months, recurrence_flag, recurrence_site, time_to_recurrence_months, vital_status, followup_duration_months, quality_of_life_eortc_qlq_c30, low_anterior_resection_syndrome

CEA Longitudinal Panel (4 cols × ~3,300 rows)

patient_id, timepoint_months (0,3,6,...,120), cea_ng_ml, assessment_type

---

Use cases

1. Molecular subtype classification — train classifiers using clinical features → MSI status, KRAS/NRAS/BRAF.
2. NCCN guideline-concordance audit — measure how often Pembrolizumab is used in MSI-H, anti-EGFR in RAS WT, BEACON-CRC in BRAF V600E, adjuvant in Stage III.
3. Survival modeling (relative) — Cox PH on OS/DFS by stage + molecular features (note: absolute survival values are shorter than literature due to a generator bug — see Limitations #1).
4. CEA trajectory modeling — longitudinal mixed-effects models on the CEA panel; predict recurrence from CEA dynamics.
5. Anti-EGFR / IO biomarker stratification — quasi-experimental analyses of treatment selection.
6. Anastomotic leak prediction — patient + surgical features → leak probability.
7. pCR prediction in rectal neoadjuvant — predict pathologic complete response from clinical + molecular features.
8. CRC mortality decomposition — Dead-CRC vs Dead-Other competing risks analyses.
9. Liquid biopsy ctDNA modeling — ctDNA detection + VAF by stage.
10. Teaching & training — oncology fellows, surgical residents, ML-for-healthcare bootcamps.

---

Loading examples

pandas (cohort + longitudinal)

import pandas as pd
df = pd.read_csv("hconc004_sample.csv")
cea = pd.read_csv("hconc004_cea_longitudinal.csv")

print(df.shape)         # (500, 105)
print(cea.shape)        # (~3,300, 4)
print(df["ajcc_stage_group"].value_counts())

# Join: cohort + CEA for trajectory analyses
merged = cea.merge(df[["patient_id", "ajcc_stage_group", "recist_best_response"]],
                   on="patient_id")

Hugging Face datasets

from datasets import load_dataset
ds = load_dataset("xpertsystems/hconc004-sample")
df = ds["train"].to_pandas()

MSI-H stratified IO benefit

mets = df[df["clinical_m_stage"].isin(["M1a","M1b","M1c"])]
io_use = mets.groupby("msi_status")["chemotherapy_regimen_line1"].apply(
    lambda s: s.isin(["Pembrolizumab", "Nivolumab+Ipilimumab"]).mean()
)
print(io_use)
# MSI-High: ~80% IO; MSS: ~0% (structural)

CEA trajectory by RECIST response

import matplotlib.pyplot as plt

merged = cea.merge(df[["patient_id","recist_best_response"]], on="patient_id")
for resp in ["CR","PR","SD","PD"]:
    sub = merged[merged["recist_best_response"] == resp]
    if len(sub) == 0: continue
    avg = sub.groupby("timepoint_months")["cea_ng_ml"].median()
    plt.plot(avg.index, avg.values, label=resp, marker='o')
plt.yscale("log")
plt.legend(); plt.xlabel("Months from baseline"); plt.ylabel("CEA (ng/mL)")
plt.title("Median CEA Trajectory by RECIST Response")
plt.show()

KRAS/BRAF mutual exclusivity check

co_mut = df[(df["kras_codon12_mutation"] != "WT") &
            (df["braf_v600e_status"] == "V600E")]
print(f"KRAS+BRAF co-occurrence: {len(co_mut)} (should be 0)")

co_mut2 = df[(df["kras_codon12_mutation"] != "WT") &
             (df["nras_mutation"] != "WT")]
print(f"KRAS+NRAS co-occurrence: {len(co_mut2)} (should be 0)")

Lymph node adequacy audit

surgical = df[df["surgery_intent"] != "None"]
adequacy = (surgical["lymph_nodes_harvested"] >= 12).mean()
print(f"NCCN LN adequacy (≥12): {adequacy:.1%} (target ≥85%)")

---

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. 🚨 SEVERE — Weibull survival sampling bug. The generator's Weibull sampling formula at lines 828-831 is incorrect:

   lam = median_arr / (np.log(2) ** (1/k))
   return (-lam * np.log(1 - u)) ** (1/k)   # BUG: lam inside power
   

   The correct inverse-CDF form is lam * (-np.log(1 - u)) ** (1/k). The bug places the scale parameter lam inside the exponentiation rather than outside, producing dramatically shortened survival times across all survival endpoints (OS, DFS, PFS).

   Observed vs target medians:

   • Stage I OS: observed ~28mo vs target ~120mo (23% of target)
   • Stage IV OS: observed ~6mo vs target ~20mo (30% of target)
   • PFS FOLFOX: observed ~4mo vs target ~10mo (40% of target)

   Relative ordering IS preserved — Stage I OS > Stage III OS > Stage IV OS monotonicity holds across all seeds. Use survival data for relative benchmarking only, not for absolute landmark survival estimates. The vital_status field correctly reflects observed-vs-followup but at shortened timescales. Scorecard OS metrics are calibrated to OBSERVED ranges to reflect generator output, with the discrepancy disclosed here. The full commercial product fixes the formula.

2. BRAF V600E in MSI-H is observed at ~10-19% vs literature ~30-40%. The generator assigns BRAF V600E at 30% probability in MSI-H, but the mutual-exclusivity override at line 351 zeros out cases where BRAF would coincide with RAS mutation. Since some MSI-H patients are RAS-mutant, they get the BRAF override applied, pulling the rate down.

3. HER2 amplification at ~5-8% vs literature ~3-5%. Slight enrichment in the RAS WT + BRAF WT subset (where HER2 amp is most common); cohort percentage trends a bit high vs published.

4. MSI-Low is over-represented at ~4-5%. The generator assigns MSI-L at 5% probability unconditionally; published MSI-L prevalence is <2%. This category is also clinically ambiguous and often grouped with MSS in modern guidelines.

5. Dead-CRC rate (78-82%) is dramatically high. Driven by the Weibull bug (#1) — most patients have their OS draw below the follow-up window (os_months <= followup), triggering death attribution. In real cohorts with 5-year follow-up, ~30-50% would be deceased.

6. CEA longitudinal panel has VARIABLE rows per patient (3-16, median 6). The panel truncates at tp > os_mo + 6 (line 910), so shorter survivors have fewer CEA visits. Cannot use this panel for fixed-N visit analyses without filtering. Join on patient_id and groupby is safe.

7. lynch = (staging.index < 0) is dead code at line 314 of the generator (placeholder never used). Lynch syndrome assignment is correctly done in demographics module via lynch_syndrome_flag.

8. Some "Anti-EGFR in non-RAS-WT" cases exist (~3 per 500) — these are the MSI-H FOLFIRI+Cetuximab branch (line 676-677), an intentional exception to the RAS WT gating because MSI-H patients can get IO- ineligible-default-to-EGFR. Not a violation, but worth knowing for filter logic.

9. recist_depth_of_response_pct for SD covers a wide range (-29 to +10), which overlaps with PR (-30 to -80) and PD (+11 to +50) by 1 percentage point at the boundaries. RECIST 1.1 actually defines PR as ≥30% decrease, PD as ≥20% increase — generator's PR is correct (-30 to -80), PD is slightly conservative (≥11% instead of ≥20%).

10. datetime.utcnow() is deprecated (line 1014) — used for metadata timestamp, harmless but emits a DeprecationWarning in modern Python. Replace with datetime.now(timezone.utc).

11. Race/ethnicity is not coupled to outcomes. Real CRC epidemiology shows substantial racial disparities (Black patients have ~20% higher CRC mortality, lower MSI-H prevalence, earlier age at diagnosis). The synthetic cohort is intentionally race-blinded in outcomes to avoid encoding disparity bias into trainees' models.

12. PFS only assigned to palliative chemo cohort. Adjuvant patients have progression_free_survival_months = NaN. For DFS-style analyses in adjuvant patients, use disease_free_survival_months.

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
Cohort patients	500	20,000+ (configurable)
CEA panel	~3,300 rows (variable)	Configurable cadence (fixed N option)
Weibull survival bug	YES (disclosed)	FIXED — literature-calibrated survival
Absolute OS	~30% of target	Matches MOSAIC/FIRE-3/KEYNOTE-177
BRAF in MSI-H	~10-15% (disclosed)	Literature 30-40%
Race-outcome coupling	None (race-blinded)	Configurable disparity profiles
Validation report	Yes (34 metrics)	Yes + custom scorecard
Format	CSV	CSV, Parquet, JSON
License	CC-BY-NC-4.0 (non-commercial)	Commercial use license
Schema mapping	—	SEER / NCCN / NCDB / TCGA-COADREAD
Treatment line 2-3	First-line only	Multi-line cascade
Support	Community	Email / SLA

---

Citation

@dataset{xpertsystems_hconc004_2026,
  title  = {HC-ONC-004: Colorectal Cancer Synthetic Cohort with CEA Longitudinal Panel},
  author = {{XpertSystems.ai}},
  year   = {2026},
  version= {1.0.0},
  url    = {https://huggingface.co/datasets/xpertsystems/hconc004-sample},
  license= {CC-BY-NC-4.0 (sample); Commercial (full product)},
  note   = {Calibrated against SEER CRC 2017-2021, TCGA COADREAD, NCCN CRC/Rectal Guidelines 2024, AJCC 8th Edition, MOSAIC (Andre 2009), FIRE-3 (Heinemann 2014), KEYNOTE-177 (Andre 2020), BEACON-CRC (Kopetz 2019), TRIBE2 (Cremolini 2020), CheckMate 142 (Overman 2018), KRYSTAL-1 (Skoulidis 2021), Engstrand 2018 (liver mets epidemiology), Hampel 2008 (Lynch syndrome).}
}

---

Contact

• Email: pradeep@xpertsystems.ai
• Web: https://xpertsystems.ai
• Vertical: Healthcare / Oncology
• SKU catalog: SKU 4 of the Oncology vertical (14 SKUs total across Cardiology + Oncology); ~79 SKUs across 8 verticals

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