xpertsystems/hconc006-sample

---
license: cc-by-nc-4.0
language:
- en
tags:
- synthetic-data
- healthcare
- oncology
- hepatocellular-carcinoma
- hcc
- liver-cancer
- bclc
- child-pugh
- meld
- afp
- tace
- milan-criteria
- liver-transplant
- imbrave150
- himalaya
- tcga-lihc
- longitudinal
- xpertsystems
pretty_name: "HC-ONC-006 — Liver Cancer (HCC) Synthetic Cohort (sample)"
size_categories:
- 1K<n<10K
task_categories:
- tabular-classification
- tabular-regression
- time-series-forecasting
- survival-analysis
---

# HC-ONC-006 — Liver Cancer (HCC) Synthetic Cohort

**Sample dataset (500-patient primary cohort + ~4,600-row AFP longitudinal panel) from the XpertSystems.ai Synthetic Data Factory — Oncology vertical, SKU 6**

A fully synthetic **hepatocellular carcinoma (HCC)** cohort spanning the
complete clinical pathway: cirrhosis etiology stratification (HBV/HCV/ALD/
NASH/Cryptogenic), comprehensive hepatic reserve assessment (Child-Pugh A/B/C,
MELD, MELD-Na, ascites grade, encephalopathy, portal hypertension, varices,
PVT), BCLC staging (0/A/B/C/D) with tumor burden detail (count, largest
diameter, sum of diameters, Milan + UCSF criteria, macrovascular invasion,
extrahepatic spread, satellite nodules, rupture, bilobar distribution),
AFP / AFP-L3 / DCP (PIVKA-II) biomarker panel with response/progression
flags and doubling time, comprehensive molecular profiling (TERT promoter
mutations, CTNNB1, TP53, ARID1A, MET amplification, FGF19, CDKN2A, PD-L1
TPS+CPS, TMB, histologic grade, Wnt pathway), locoregional therapy
(TACE-DEB/conventional, TARE-Y90 with dosimetry, RFA/MWA/PEI/Cryoablation
with mRECIST response, downstaging, bridge therapy), liver transplant
evaluation (Milan/UCSF/Downstaged/Metroticket criteria, MELD exception,
waitlist, donor type LDLT/DDLT, ischemia times, post-transplant outcomes
including recurrence with site and timing, rejection, immunosuppression),
surgical resection (Open/Laparoscopic/Robotic, R-status, FLR, EBL, Pringle,
operative time, post-hepatectomy liver failure, bile leak, LOS, RFS),
**IMbrave150/HIMALAYA-era systemic therapy** (Atezolizumab+Bevacizumab,
Durvalumab+Tremelimumab, Sorafenib, Lenvatinib, Durvalumab; second-line
Cabozantinib/Regorafenib/Ramucirumab/Pembrolizumab) with RECIST response,
PFS/OS by regimen calibrated to landmark trials, comprehensive toxicity
profiling (sorafenib HFSR, bevacizumab HTN/bleeding, irAE type and grade,
lenvatinib dose reduction), survival endpoints, QoL (FACT-Hep, EQ-5D),
and a **quarterly AFP+DCP longitudinal panel** 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-006 |
| **Vertical** | Healthcare → Oncology (SKU 6) |
| **Tables** | 2 (primary + AFP/DCP longitudinal) |
| **Sample size** | 500-patient primary × 126 columns; ~4,600-row AFP panel × 5 cols |
| **Follow-up** | Quarterly AFP/DCP for up to 40 quarters (variable per patient — truncated by OS) |
| **Standards** | BCLC Staging Schema, AJCC 8th Edition, Child-Pugh, MELD/MELD-Na, Milan Criteria, UCSF Criteria, mRECIST, RECIST 1.1, ISGPF/ISGPS |
| **Format** | CSV (2 tables) |
| **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

HCC data is uniquely fragmented: SEER provides population-level data but
no liver-function or molecular detail; TCGA LIHC has deep genomics but
n=377; landmark clinical trials (IMbrave150, HIMALAYA, SHARP, REFLECT,
CELESTIAL, REACH-2) are restricted; transplant registries (UNOS, ELTR)
require special access. This synthetic cohort gives you the **full HCC
phenome with liver function, BCLC staging, biomarkers, transplant pathway,
and IMbrave150-era systemic therapy** in two relational tables joined on
`patient_id`:

- ✅ **Etiology ↔ HBV/HCV serology coupling** — HBsAg+ iff HBV etiology
(0 leak); HCV genotype only set when HCV; SVR rate ~80% (DAA era mix)
- ✅ **BCLC ↔ treatment gating** — Resection ⊂ BCLC-0/A ∩ CP-A ∩ ¬listed
(NCCN concordance); Transplant ⊂ Milan-met ∪ downstaged; Systemic ⊂ ¬CP-C;
BCLC-D excludes all curative therapy (0 violations)
- ✅ **Child-Pugh + MELD calculated from underlying labs** — bilirubin,
albumin, INR, creatinine, sodium drive UNOS formula with ascites +
encephalopathy points
- ✅ **TERT/CTNNB1/TP53/ARID1A frequencies match TCGA LIHC** — TERT 56-61%,
CTNNB1 22-29%, TP53 37-44%, ARID1A 13-18%
- ✅ **Milan + UCSF criteria computed structurally** from tumor count +
largest diameter; downstaging updates Milan status post-LRT
- ✅ **IMbrave150-era treatment uptake** — Atezo+Bev ~38-46% first-line,
Durva+Treme ~15-22%, Sorafenib ~17-28% (mixed era)
- ✅ **Trial-anchored OS/PFS by regimen** — IMbrave150 (atezo+bev OS ~19mo),
HIMALAYA (durva+treme ~16mo), SHARP (sorafenib ~11mo), REFLECT
(lenvatinib ~14mo)
- ✅ **REACH-2 ramucirumab restricted to AFP≥400** (0 violations)
- ✅ **Transplant outcomes** — Milan/UCSF/Downstaged listing, MELD
exception, LDLT vs DDLT by region (Asia-Pacific LDLT 60%), post-transplant
recurrence ~16%

Coverage spans:
- **AJCC 8th Edition + BCLC staging** (Very Early through Terminal)
- **Cirrhosis etiology** — HBV / HCV / ALD / NASH / Cryptogenic with
HBV-DNA, HBV genotype (A/B/C/D), HCV genotype (1a/1b/2/3/4), SVR flag
- **Hepatic reserve** — bilirubin, albumin, INR, creatinine, sodium,
ascites grade, encephalopathy 0-4, Child-Pugh score + class, MELD, MELD-Na,
portal hypertension, esophageal varices grade, PVT, liver stiffness
(FibroScan), fibrosis stage (F3/F4)
- **Tumor burden** — count, largest diameter, sum of diameters, Milan +
UCSF, macrovascular invasion, extrahepatic spread, satellite nodules,
rupture, bilobar
- **Biomarkers** — AFP baseline + nadir + doubling time + response +
progression, AFP-L3 fraction, DCP/PIVKA-II, Ramucirumab eligibility
(AFP≥400)
- **Molecular** — TERT promoter (C228T/C250T), CTNNB1, TP53, ARID1A,
MET amplification, FGF19 amplification, CDKN2A deletion, PD-L1 TPS+CPS,
TMB, histologic grade G1-G4, Wnt pathway, AFP model score
- **Locoregional therapy** — TACE-DEB / TACE conventional / TARE-Y90 /
RFA / MWA / PEI / Cryoablation with mRECIST response, TACE sessions +
embolization agent, Y90 dosimetry + lung shunt, downstaging,
Milan-post-LRT, bridge therapy, complications, transition to systemic
- **Transplant** — evaluation, listing with criteria, MELD exception,
waitlist time, dropout reason, donor type, ischemia times, post-transplant
recurrence + site + timing, acute/chronic rejection, immunosuppression
(Tacrolimus/Cyclosporine/mTOR), 5-year graft survival
- **Surgical resection** — Laparoscopic/Open/Robotic, Anatomic/Non-anatomic,
R0/R1/R2, future liver remnant, EBL, Pringle, operative time,
post-hepatectomy liver failure, bile leak, LOS, RFS
- **Systemic therapy** — Atezo+Bev / Durva+Treme / Sorafenib / Lenvatinib /
Durvalumab first-line; Cabozantinib / Regorafenib / Ramucirumab /
Pembrolizumab / BSC second-line; RECIST response, depth of response,
PFS, OS by regimen
- **Toxicities** — sorafenib HFSR grade, bevacizumab HTN + bleeding,
irAE type (Hepatitis/Colitis/Dermatitis/Thyroiditis/Pneumonitis) + grade,
lenvatinib dose reduction
- **Outcomes** — overall survival, vital status (Alive/Dead-HCC/Dead-Other),
FACT-Hep score, EQ-5D utility
- **AFP longitudinal panel** — quarterly AFP + DCP for up to 40 quarters
(truncated by OS)

---

## Calibration anchors (industry-grade)

This cohort is calibrated against named registries, guidelines, and trials.
Selection from the 43-metric scorecard:

| Metric | Sample value (seed 42) | Target range | Source |
|---|---:|---|---|
| Mean age | 62.9 yr | 58–68 | SEER HCC ~64 |
| Male % | 74.0% | 65–82 | SEER HCC ~70-75% |
| HBV etiology | 31.8% | 25–38 | Global HCC ~30-40% |
| HCV etiology | 26.6% | 18–33 | Global HCC ~20-30% |
| NASH etiology | 17.4% | 12–25 | Rising etiology ~15-25% |
| Child-Pugh A | 22.8% | 15–32 | Cohort under-represents (disclosed) |
| Child-Pugh C | 10.6% | 5–16 | HCC at-dx ~10-15% |
| MELD median | 11.7 | 9–14 | HCC MELD ~10-14 |
| PVT % | 10.0% | 7–18 | HCC PVT ~10-15% |
| BCLC-A % | 33.0% | 28–43 | Cohort design ~35% |
| BCLC-C % | 28.2% | 24–36 | Cohort design ~30% |
| AFP ≥400 ng/mL | 14.8% | 8–22 | Cohort distribution |
| Milan criteria met | 16.0% | 12–25 | Cohort tumor-burden-driven |
| TERT promoter mutation | 61.2% | 50–65 | TCGA LIHC ~55-60% |
| CTNNB1 mutation | 25.4% | 18–32 | TCGA LIHC ~25% |
| TP53 mutation | 44.0% | 32–48 | TCGA LIHC ~30-40% |
| ARID1A mutation | 18.0% | 10–22 | TCGA LIHC ~16% |
| HBsAg+ in HBV | 100% | ≥100% (floor) | Structural |
| SVR in HCV | 82.0% | 70–90 | DAA era ~95%; cohort mixed |
| LRT overall | 49.4% | 38–56 | NCCN LRT ~40-55% |
| TACE in BCLC-B | 29.4% | 22–42 | Generator ~30%; clinical ~75% (disclosed) |
| Transplant evaluation | 30.0% | 25–38 | Cohort design |
| Transplant listed | 10.8% | 5–14 | HCC listing ~10% |
| Transplant performed | 6.8% | 3–10 | Cohort transplanted ~5-7% |
| Systemic in BCLC-C | 79.4% | 65–90 | Eligible cohort |
| Atezo+Bev in systemic | 38.5% | 25–50 | IMbrave150 era |
| Durva+Treme in systemic | 22.4% | 12–28 | HIMALAYA era |
| ORR atezo+bev | 38.2% | 18–42 | IMbrave150 ORR 27.3% |
| OS median overall | 18.85 mo | 15–25 | Mixed cohort |
| OS median BCLC-0/A | 40.0 mo | 28–60 | Curative cohort |
| OS median BCLC-C | 8.7 mo | 6–14 | Pre-IMbrave era ~10mo |
| OS median BCLC-D | 2.2 mo | 1.5–6 | Terminal stage |
| Transplant only if listed | 100% | ≥100% (floor) | Structural |
| Systemic excludes CP-C | 100% | ≥100% (floor) | Structural |
| Ramucirumab only AFP≥400 | 100% | ≥100% (floor) | REACH-2 |
| BCLC-D no curative | 100% | ≥100% (floor) | Structural |
| Stage OS monotonic | 100% | ≥100% (floor) | Structural |

Full 43-metric scorecard ships in `validation_report.json` and `validation_report.md`.

---

## Files in this sample

```
hconc006_sample/
├── hconc006_sample.csv # 500 patients × 126 columns (primary)
├── hconc006_afp_longitudinal.csv # ~4,600 rows × 5 cols (AFP+DCP 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
```

**Both tables join on `patient_id`.** The AFP longitudinal table has
**variable rows per patient** (median 7, range 1-40) — quarterly
visits truncated at `min(OS_months / 3 + 1, 40)`. Columns:
`patient_id, timepoint_quarter, timepoint_months, afp_ng_ml, dcp_pivka2_mau_ml`.

---

## Schema (126 columns in primary cohort across 10 modules)

### Primary: Demographics (8 cols)
`patient_id`, `age_at_diagnosis`, `sex`, `race`, `geographic_region`,
`cirrhosis_etiology`, `diagnosis_year`, `diagnosis_month`

### Primary: Hepatic Reserve (16 cols)
`serum_bilirubin_mg_dl`, `serum_albumin_g_dl`, `inr`,
`serum_creatinine_mg_dl`, `serum_sodium_meq_l`, `ascites_grade`,
`hepatic_encephalopathy_grade`, `child_pugh_score`, `child_pugh_class`,
`meld_score`, `meld_na_score`, `portal_hypertension_flag`,
`esophageal_varices_grade`, `portal_vein_thrombosis_flag`,
`liver_stiffness_kpa`, `fibrosis_stage_metavir`

### Primary: BCLC Staging (12 cols)
`bclc_stage`, `tumor_count`, `largest_tumor_diameter_cm`,
`sum_of_diameters_cm`, `milan_criteria_met_flag`, `ucsf_criteria_met_flag`,
`macrovascular_invasion_flag`, `extrahepatic_spread_flag`, `ecog_ps`,
`satellite_nodules_flag`, `tumor_rupture_flag`, `bilobar_distribution_flag`

### Primary: AFP Biomarkers (8 cols)
`afp_at_baseline_ng_ml`, `afp_l3_fraction_pct`, `dcp_pivka2_mau_ml`,
`afp_nadir_ng_ml`, `afp_response_flag`, `afp_progression_flag`,
`afp_doubling_time_days`, `ramucirumab_eligibility_flag`

### Primary: Molecular Markers (19 cols)
`hbsag_flag`, `hbv_dna_iu_ml`, `hbv_genotype`, `hcv_rna_flag`,
`hcv_genotype`, `svr_achieved_flag`, `tert_promoter_mutation`,
`ctnnb1_mutation`, `tp53_mutation`, `arid1a_mutation`,
`met_amplification`, `fgf19_amplification`, `cdkn2a_deletion`,
`pd_l1_tumor_proportion_score`, `pd_l1_combined_positive_score`,
`tmb_mutations_per_mb`, `histologic_grade`, `wnt_pathway_activation`,
`afp_model_score`

### Primary: Locoregional Therapy (13 cols)
`lrt_performed_flag`, `lrt_modality`, `mrecist_response_lrt`,
`tace_sessions_count`, `tace_embolization_agent`, `y90_dosimetry_gy`,
`y90_lung_shunt_fraction_pct`, `ablation_complete_response_flag`,
`downstaging_success_flag`, `milan_criteria_post_lrt_flag`,
`waitlist_bridge_therapy_flag`, `lrt_complication_flag`,
`lrt_to_systemic_transition_flag`

### Primary: Transplant (19 cols)
`transplant_evaluation_flag`, `transplant_listed_flag`,
`transplant_listing_criteria`, `meld_exception_points_granted`,
`meld_exception_score`, `waitlist_time_months`, `waitlist_dropout_reason`,
`transplant_performed_flag`, `donor_type`, `cold_ischemia_time_hours`,
`warm_ischemia_time_minutes`, `post_transplant_recurrence_flag`,
`recurrence_site_post_transplant`, `time_to_recurrence_post_transplant_months`,
`acute_rejection_flag`, `chronic_rejection_flag`, `immunosuppression_protocol`,
`mtor_inhibitor_flag`, `graft_survival_5yr_flag`

### Primary: Surgical Resection (12 cols)
`resection_performed_flag`, `resection_approach`, `extent_of_resection`,
`r_status`, `future_liver_remnant_pct`, `intraoperative_blood_loss_ml`,
`pringle_maneuver_flag`, `operative_time_minutes`,
`post_hepatectomy_liver_failure_flag`, `bile_leak_flag`,
`hospital_los_days`, `rfs_months_post_resection`

### Primary: Systemic Therapy (15 cols)
`systemic_therapy_flag`, `first_line_regimen`, `recist_response_first_line`,
`depth_of_response_pct`, `pfs_first_line_months`, `os_first_line_months`,
`sorafenib_hfsr_grade`, `bevacizumab_hypertension_flag`,
`bevacizumab_bleeding_event_flag`, `immune_checkpoint_irae_flag`,
`irae_type`, `irae_grade`, `lenvatinib_dose_reduction_flag`,
`second_line_regimen`, `os_second_line_months`

### Primary: Outcomes (4 cols)
`overall_survival_months`, `vital_status`, `qol_fact_hep_score`,
`qol_eq5d_utility`

### AFP Longitudinal Panel (5 cols × ~4,600 rows)
`patient_id`, `timepoint_quarter` (0, 1, 2, ..., 39),
`timepoint_months` (0, 3, 6, ..., 117), `afp_ng_ml`, `dcp_pivka2_mau_ml`

---

## Use cases

1. **BCLC-stratified survival modeling** — Cox PH on OS by BCLC stage +
liver function + molecular features.
2. **Transplant eligibility prediction** — predict Milan/UCSF criteria
met from tumor + AFP + molecular features.
3. **NCCN/EASL guideline-concordance audit** — measure how often resection
used in BCLC-0/A CP-A, transplant in Milan-met, systemic in BCLC-C.
4. **AFP trajectory modeling** — longitudinal mixed-effects models on the
AFP+DCP panel; predict recurrence from kinetics.
5. **TACE response prediction** — predict mRECIST CR/PR from patient +
tumor + LRT modality features.
6. **IMbrave150-era treatment effectiveness** — quasi-experimental
atezo+bev vs sorafenib OS comparisons.
7. **PHLF risk prediction post-resection** — predict post-hepatectomy
liver failure from preop features.
8. **Transplant waitlist mortality** — competing risks: transplant vs
tumor progression vs death.
9. **Etiology-stratified molecular subtyping** — HBV-CTNNB1 vs HCV-TERT
subtype enrichment.
10. **Teaching & training** — hepatology fellows, transplant surgery
residents, ML-for-healthcare bootcamps.

---

## Loading examples

### pandas (both tables)
```python
import pandas as pd
df = pd.read_csv("hconc006_sample.csv")
afp = pd.read_csv("hconc006_afp_longitudinal.csv")

print(df.shape) # (500, 126)
print(afp.shape) # (~4600, 5)
print(df["bclc_stage"].value_counts().sort_index())

# Join cohort + AFP for trajectory analyses
merged = afp.merge(df[["patient_id", "bclc_stage", "afp_response_flag"]],
on="patient_id")
```

### Hugging Face `datasets`
```python
from datasets import load_dataset
ds = load_dataset("xpertsystems/hconc006-sample")
df = ds["train"].to_pandas()
```

### BCLC-stratified survival curves
```python
from lifelines import KaplanMeierFitter
import matplotlib.pyplot as plt

df["dead"] = (df["vital_status"] != "Alive").astype(int)
kmf = KaplanMeierFitter()
for stage in ["BCLC-0", "BCLC-A", "BCLC-B", "BCLC-C", "BCLC-D"]:
sub = df[df["bclc_stage"] == stage]
if len(sub) < 5: continue
kmf.fit(sub["overall_survival_months"], event_observed=sub["dead"], label=stage)
kmf.plot_survival_function()
plt.title("OS by BCLC Stage in HCC")
plt.xlabel("Months"); plt.ylabel("Survival probability")
plt.show()
```

### AFP trajectory by response flag
```python
import matplotlib.pyplot as plt

merged = afp.merge(df[["patient_id", "afp_response_flag"]], on="patient_id")
for resp_label, resp_val in [("Responder", 1), ("Non-responder", 0)]:
sub = merged[merged["afp_response_flag"] == resp_val]
avg = sub.groupby("timepoint_months")["afp_ng_ml"].median()
plt.plot(avg.index, avg.values, label=resp_label, marker='o', markersize=3)
plt.yscale("log")
plt.xlabel("Months from diagnosis"); plt.ylabel("AFP (ng/mL, log scale)")
plt.legend(); plt.title("Median AFP Trajectory by Response")
plt.show()
```

### IMbrave150-era treatment comparison
```python
systemic = df[df["systemic_therapy_flag"] == 1]
treatment_os = systemic.groupby("first_line_regimen").agg(
n=("patient_id", "count"),
os_median=("os_first_line_months", "median"),
orr=("recist_response_first_line", lambda s: s.isin(["CR","PR"]).mean())
).round(2)
print(treatment_os)
```

### Milan criteria + transplant pathway
```python
milan_met = df[df["milan_criteria_met_flag"] == 1]
listing_rate = milan_met["transplant_listed_flag"].mean()
transplant_rate = milan_met["transplant_performed_flag"].mean()
print(f"Milan-met patients: {len(milan_met)} ({len(milan_met)/len(df):.1%})")
print(f" Listed for transplant: {listing_rate:.1%}")
print(f" Received transplant: {transplant_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. **Cohort skews toward worse hepatic function.** Child-Pugh A is only
~24% of this synthetic cohort vs ~50-60% in real-world HCC at diagnosis.
The generator's lab distributions (bilirubin lognormal, albumin normal
3.4 ± 0.5, INR lognormal) plus ascites/encephalopathy assignment
produce a cohort enriched in CP-B (~63%) and CP-C (~10%). **For
modeling early-stage curative-intent populations, filter the dataset
to `child_pugh_class == "A"` before training.** The full commercial
product offers configurable hepatic reserve profiles.

2. **TACE in BCLC-B is under-represented (~30% vs ~75% real-world).** The
generator's locoregional therapy assignment chooses TACE-DEB/conventional
only ~50% of the time for LRT-performed patients regardless of BCLC
stage. In clinical practice, BCLC-B is overwhelmingly TACE-treated
first-line (NCCN/EASL Class I). **For TACE-cohort analyses, expect
30-35% TACE uptake in BCLC-B; the full product offers BCLC-aware LRT
assignment.**

3. **AFP ≥400 ng/mL is under-represented (~13% vs ~30-40% literature).**
The generator's AFP distribution centers around log-normal(4.8, 1.8)
for BCLC-C (median ~120 ng/mL), so the 400 threshold is hit less often.
`ramucirumab_eligibility_flag` is structurally correct but the eligible
population is smaller than published REACH-2 / clinical screening rates.
This propagates to ramucirumab uptake in the synthetic cohort.

4. **SVR in HCV cohort ~80% (vs DAA-era ~95%).** The generator uses a
80% SVR probability uniformly across HCV patients, which reflects a
mixed-era cohort (some pre-DAA, some DAA-era). For DAA-era-only
modeling, filter on `diagnosis_year >= 2018` and treat SVR as ~95%
in your downstream analyses.

5. **`generate_locoregional_therapy` contains a dead code branch**
(line 385): `df["bclc_stage"].values.isin(...)` — `.values` returns
a numpy array which has no `.isin()` method. A `hasattr` check on
the same line catches the issue and falls through to a list-
comprehension alternative, so the function works correctly. **Bug is
silent at runtime but would crash if the `hasattr` short-circuit were
ever removed.** The full product cleans this up.

6. **`overall_survival_months` is BCLC-driven, not regimen-driven.**
The `generate_outcomes` module overwrites the regimen-specific
`os_first_line_months` (calibrated to IMbrave150/SHARP/REFLECT) with
a BCLC-stage Weibull draw. As a result, two patients on the same
regimen with the same BCLC stage will have OS distributions reflecting
stage epidemiology, not regimen efficacy. **For regimen-specific
survival modeling, use `os_first_line_months` field instead.** The
`overall_survival_months` is appropriate for cohort-level / staging-
based analyses.

7. **`overall_survival_months` for transplant patients is independently
drawn** at line 740 (Weibull k=2.0, scale=72), overriding the BCLC
draw. This produces ~72-month median OS for transplanted patients,
which is plausible but doesn't account for individual recipient
characteristics (MELD, donor type, recurrence).

8. **Post-transplant recurrence is conditioned on `transplant_performed_flag == 1`
AND p=0.16 uniform** — does NOT depend on Milan vs UCSF vs Downstaged
listing criteria. Real-world data shows downstaged patients have
higher recurrence rates than Milan-standard. The full product offers
listing-criteria-stratified recurrence.

9. **Macrovascular invasion in BCLC-C is ~25% (vs ~30-50% real-world).**
Generator assigns `rng.uniform() < 0.25` for BCLC-C macroinvasion;
clinical literature places this 30-50%. Slight under-representation.

10. **`bilobar_distribution_flag` formula at line 211 looks unusual** —
`(rng.uniform(0, 1, n) < 0.30 * (tumor_count > 1)).astype(int)` —
but works correctly via numpy broadcasting. Bilobar disease is
structurally restricted to multi-tumor patients (only when
`tumor_count > 1`) with 30% probability — clinically appropriate.

11. **Race not coupled to etiology.** Real HCC epidemiology shows
Asian-Pacific has higher HBV prevalence (~50-60% of HCC), while
European/American HCC is more HCV/NASH-driven. Cohort intentionally
decouples for race-blind outcome modeling.

12. **`afp_l3_fraction_pct` is uniform [5, 90]** — not coupled to AFP
baseline or BCLC stage. Real AFP-L3 fraction correlates strongly
with HCC vs benign liver disease. For AFP-L3 modeling, treat as
a noise channel.

13. **`overall_survival_months` capped at 180 months** for non-
transplanted patients (and 180 for transplanted). Very-Early stage
survivors are right-censored at 15-year horizon.

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) |
| AFP panel | ~4,600 rows (variable) | Configurable cadence |
| Child-Pugh A enrichment | ~24% (disclosed) | Configurable 50-70% |
| TACE in BCLC-B | ~30% (disclosed) | NCCN-concordant ~75% |
| AFP ≥400 | ~13% | Configurable 25-40% |
| SVR in HCV | ~80% mixed era | Configurable by dx year |
| OS regimen-driven | BCLC-overwrite (disclosed) | Regimen-preserving option |
| Race-etiology coupling | None | Configurable by region |
| Validation report | Yes (43 metrics) | Yes + custom scorecard |
| Format | CSV | CSV, Parquet, JSON |
| License | CC-BY-NC-4.0 (non-commercial) | Commercial use license |
| Schema mapping | — | SEER / NCCN / UNOS / TCGA-LIHC |
| Multi-line treatment | First + Second | Multi-line cascade |
| Support | Community | Email / SLA |

---

## Citation

```bibtex
@dataset{xpertsystems_hconc006_2026,
title = {HC-ONC-006: Liver Cancer (HCC) Synthetic Cohort with BCLC Staging, Transplant Pathway, and IMbrave150-Era Systemic Therapy},
author = {{XpertSystems.ai}},
year = {2026},
version= {1.0.0},
url = {https://huggingface.co/datasets/xpertsystems/hconc006-sample},
license= {CC-BY-NC-4.0 (sample); Commercial (full product)},
note = {Calibrated against SEER HCC 2017-2021, TCGA LIHC molecular frequencies (Cancer Genome Atlas Research Network 2017), BCLC Staging Schema 2022 (Reig 2022), AJCC 8th Edition, Child-Pugh Score, MELD (Kamath 2001), Milan Criteria (Mazzaferro 1996), UCSF Criteria (Yao 2001), IMbrave150 (Finn 2020 atezolizumab+bevacizumab), HIMALAYA (Abou-Alfa 2022 durvalumab+tremelimumab), SHARP (Llovet 2008 sorafenib), REFLECT (Kudo 2018 lenvatinib), CELESTIAL (Abou-Alfa 2018 cabozantinib), REACH-2 (Zhu 2019 ramucirumab AFP≥400).}
}
```

---

## Contact

- **Email:** [[e-mail]](mailto:[e-mail])
- **Web:** [https://xpertsystems.ai](https://xpertsystems.ai)
- **Vertical:** Healthcare / Oncology
- **SKU catalog:** SKU 6 of the Oncology vertical (16 SKUs total across Cardiology + Oncology); ~81 SKUs across 8 verticals

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