--- license: cc-by-nc-4.0 language: - en tags: - synthetic-data - healthcare - oncology - pancreatic-cancer - pdac - kras - smad4 - ca19-9 - folfirinox - olaparib - prodige - mpact - polo - tcga-paad - longitudinal - multi-table - xpertsystems pretty_name: "HC-ONC-005 — Pancreatic Cancer (PDAC) Synthetic Cohort (sample)" size_categories: - 1K8) - **Tumor anatomy** — head/body/tail subsite, tumor size, jaundice + biliary stent flags - **KRAS molecular detail** — 7 codon variants + WT, allelic frequency, G12C targeted-therapy flag - **Tumor suppressors** — TP53, SMAD4, CDKN2A (Iacobuzio-Donahue 4-gene set) - **HRR/DDR panel** — BRCA1/2 germline + BRCA2 somatic, PALB2, ATM, HRD score, PARP eligibility composite - **Immune biomarkers** — MSI, TMB, PD-L1 TPS, HER2, NTRK fusion - **Liquid biopsy** — CA 19-9 baseline + Lewis-negative flag, CEA baseline, ctDNA detection + VAF, KRAS-ctDNA flag - **Surgical pathology** — Whipple/PPPD/Distal/Total/Palliative_Bypass procedures with Open/MIS/Robotic approach, R-status, LN harvest + positive + ratio, ISGPF fistula (None/A/B/C), ISGPS DGE (None/A/B/C), PPH grading, bile leak, wound infection, portal vein resection, arterial resection, operative time, EBL, ICU, LOS, 90-day readmission, pCR + nCR flags - **First-line treatment** — FOLFIRINOX/mFOLFIRINOX/Gem+NabP/NALIRIFOX/ Gem mono/Pembro (MSI-H)/BSC with cycle counts, dose reduction, toxicity, G3/4 events, febrile neutropenia, peripheral neuropathy grading - **Second-line treatment** — Olaparib (PARP-eligible)/FOLFOX/Capecitabine/ Regorafenib/BSC - **Targeted therapy** — Sotorasib/Adagrasib (G12C), Olaparib (HRR+), Pembrolizumab (MSI-H) - **Radiation** — SBRT/Conventional_EBRT/IMRT/Proton with dose - **RECIST response** — CR/PR/SD/PD with depth-of-response %, CA 19-9 response/progression flags, CA 19-9 nadir + timing - **Survival** — OS, PFS, RFS (surgical patients only), time-to-next- treatment, vital status (Alive/Dead_PDAC/Dead_Other) - **QoL** — EORTC QLQ-C30, EORTC PAN26 (PDAC-specific module) - **Longitudinal CA 19-9 panel** — monthly values from diagnosis through min(60, OS+1) months with realistic trajectory shapes --- ## Calibration anchors (industry-grade) This cohort is calibrated against named registries, guidelines, and trials. Selection from the 35-metric scorecard: | Metric | Sample value (seed 42) | Target range | Source | |---|---:|---|---| | Mean age | 70.7 yr | 66–75 | SEER PDAC median 70-71 | | Female % | 45.8% | 38–52 | SEER ~45-48% | | Diabetes % | 41.2% | 28–48 | Aggarwal 2013 | | Stage IV % | 47.0% | 38–58 | SEER ~50-55% Stage IV at dx | | Stage I % | 5.2% | 2–9 | SEER ~5% Stage I (mostly late dx) | | Head tumor % | 71.4% | 60–80 | Hidalgo 2010 | | Jaundice in head | 65.0% | 60–80 | Head PDAC ~70-80% | | KRAS mutation | 93.6% | ≥90% (floor) | TCGA PAAD ~93% | | KRAS G12D | 30.4% | 28–42 | TCGA PAAD ~36% | | KRAS G12V | 24.6% | 15–28 | TCGA PAAD ~22% | | TP53 mutation | 73.2% | 65–82 | TCGA PAAD ~70-75% | | SMAD4 lost | 48.2% | 38–55 | TCGA PAAD ~50% | | CDKN2A lost | 63.2% | 52–68 | TCGA PAAD ~60% | | BRCA2 germline | 2.8% | 1.5–8 | POLO PDAC ~5% | | MSI-H | 0.8% | 0.1–4 | Hu 2018 PDAC ~1% | | Surgery rate | 20.6% | 12–30 | Cohort Stage IV-driven | | Whipple in head surgical | 88.6% | ≥80% (floor) | Standard for head PDAC | | R0 resection | 72.8% | 60–85 | Strobel 2017 | | Fistula B/C (ISGPF) | 25.2% | 15–32 | ISGPF ~20-25% | | DGE B/C (ISGPS) | 20.4% | 12–32 | ISGPS ~15-25% | | FOLFIRINOX in palliative | 38.6% | 25–50 | PRODIGE-4 era | | Gem+NabP in palliative | 43.0% | 35–55 | MPACT era | | Pembrolizumab in MSI-H | 75.0% | ≥40% (floor) | KEYNOTE-158 | | Olaparib in PARP-eligible | 28.3% | 15–50 | POLO era | | OS median (overall) | 5.75 mo | 4.5–9 | PDAC ~6-11mo; IV-heavy cohort | | OS median Stage IV | 5.25 mo | 3.5–8 | PRODIGE-4/MPACT ~6-8mo | | PFS median | 2.4 mo | 1.5–4 | mPDAC ~3-6mo | | ORR (overall) | 22.6% | 15–35 | Mixed regimen ~20-30% | | ctDNA in Stage IV | 89.4% | 70–95 | Cohen 2018 ~80-90% | | Liver mets in Stage IV | 48.9% | 40–65 | Yachida 2010 ~50% | | Stage→Resectability monotonic | 100% | ≥100% (floor) | Structural | Full 35-metric scorecard ships in `validation_report.json` and `validation_report.md`. --- ## Files in this sample ``` hconc005_sample/ ├── hconc005_sample.csv # 500 patients × 107 columns (primary) ├── hconc005_ca19_9_longitudinal.csv # ~4,400 rows × 3 cols (CA 19-9 panel) ├── hconc005_surgical.csv # ~100 rows × 21 cols (surgical subset) ├── hconc005_molecular.csv # 500 rows × 25 cols (molecular panel) ├── hconc005_treatment_history.csv # ~860 rows × 8 cols (long-form treatment) ├── 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 ``` **All 5 tables join on `patient_id`.** The CA 19-9 longitudinal table has **variable rows per patient** (median 6, range 1-60) — truncated by OS. Columns: `patient_id, month_from_diagnosis, ca19_9_u_ml`. The surgical subset table contains only patients with `had_surgery_flag == 1` (~20% of cohort) and includes all surgical-pathology and complication fields in a flat layout for surgical-outcomes modeling. --- ## Schema (107 columns in primary cohort across 6 modules) ### Primary: Demographics (16 cols) `patient_id`, `diagnosis_date`, `age_at_diagnosis`, `sex`, `race`, `smoking_history`, `diabetes_mellitus_flag`, `family_history_flag`, `chronic_pancreatitis_flag`, `bmi_at_diagnosis`, `albumin_g_dl`, `weight_loss_6mo_kg`, `ecog_performance_status`, `pain_score_vas`, `jaundice_flag`, `biliary_stent_flag` ### Primary: Staging (17 cols) `ajcc_stage_group`, `pdac_stage_simple`, `resectability_status`, `clinical_t_stage`, `clinical_n_stage`, `clinical_m_stage`, `tumor_site`, `tumor_size_cm`, `sma_contact_degrees`, `sma_involvement_flag`, `smv_pv_involvement`, `celiac_axis_involvement_flag`, `hepatic_artery_involvement_flag`, `aorta_involvement_flag`, `liver_metastasis_flag`, `liver_metastasis_count`, `peritoneal_metastasis_flag`, `lung_metastasis_flag`, `distant_nodal_metastasis_flag` ### Primary: Molecular (24 cols) `kras_mutation` (WT/G12D/G12V/G12R/G12C/G12A/G12S/Q61H), `kras_allelic_frequency_pct`, `kras_g12c_targeted_flag`, `tp53_mutation`, `smad4_loss`, `cdkn2a_loss`, `brca1_germline`, `brca2_germline`, `brca2_somatic`, `palb2_germline`, `atm_germline`, `hrd_score`, `parp_inhibitor_eligible_flag`, `msi_status`, `tmb_mutations_per_mb`, `pdl1_tumor_proportion_score`, `her2_status`, `ntrk_fusion_flag`, `ca19_9_baseline_u_ml`, `ca19_9_lewis_negative_flag`, `cea_baseline_ng_ml`, `ctdna_detected_flag`, `ctdna_vaf_pct`, `kras_ctdna_detected_flag` ### Primary: Surgery (22 cols) `had_surgery_flag`, `neoadjuvant_therapy_flag`, `surgery_procedure`, `surgical_approach`, `r_status`, `lymph_nodes_harvested`, `lymph_nodes_positive`, `lymph_node_ratio`, `pancreatic_fistula_isgpf`, `delayed_gastric_emptying_isgps`, `post_pancreatectomy_hemorrhage`, `bile_leak_flag`, `wound_infection_flag`, `portal_vein_resection_flag`, `arterial_resection_flag`, `operative_time_minutes`, `estimated_blood_loss_ml`, `icu_admission_flag`, `hospital_los_days`, `readmission_90d_flag`, `pathologic_complete_response_flag`, `near_complete_response_flag` ### Primary: Treatment (17 cols) `first_line_regimen`, `second_line_regimen`, `targeted_agent`, `recist_response`, `recist_depth_of_response_pct`, `ca19_9_response_flag`, `ca19_9_progression_flag`, `ca19_9_nadir_u_ml`, `ca19_9_nadir_timing_weeks`, `cycles_completed`, `dose_reduction_flag`, `treatment_discontinuation_reason`, `grade3_4_toxicity_flag`, `febrile_neutropenia_flag`, `peripheral_neuropathy_grade`, `parp_response_flag`, `radiation_flag`, `radiation_type`, `radiation_dose_gy` ### Primary: Outcomes (7 cols) `overall_survival_months`, `progression_free_survival_months`, `recurrence_free_survival_months`, `time_to_next_treatment_months`, `vital_status`, `quality_of_life_qlq_c30`, `quality_of_life_pan26` ### CA 19-9 Longitudinal Panel (3 cols × ~4,400 rows) `patient_id`, `month_from_diagnosis` (0, 1, 2, ..., min(60, OS+1)), `ca19_9_u_ml` ### Surgical Subset (21 cols × ~100 rows) Joins all surgical-pathology and complication fields for the ~20% of patients who had surgery; flat layout for surgical modeling. ### Molecular Panel (25 cols × 500 rows) Flat molecular profile for the full cohort. ### Treatment History (8 cols × ~860 rows) Long-form: one row per (`patient_id`, `treatment_line`) where `treatment_line ∈ {First, Second}`. Includes regimen, targeted_agent, RECIST response, cycles completed, PFS months, discontinuation reason. --- ## Use cases 1. **PDAC molecular subtyping** — KRAS variant classification, HRD-high vs -low stratification, MSI-H detection workflows. 2. **Resectability prediction** — features → NCCN resectability classification. 3. **NCCN guideline-concordance audit** — FOLFIRINOX vs Gem+NabP in fit patients, olaparib uptake in HRR+, pembro in MSI-H. 4. **Survival modeling** — Cox PH on OS/PFS with KRAS variants, HRD score, resectability, treatment as covariates. 5. **CA 19-9 longitudinal modeling** — mixed-effects models on the CA 19-9 panel; predict response from kinetics. 6. **Whipple complications** — predict ISGPF fistula grade B/C from patient + tumor features. 7. **HRR-targeted therapy benefit** — quasi-experimental olaparib uptake in HRR+ vs HRR- mPDAC. 8. **Multi-table joins for ML** — combine primary + CA 19-9 + surgical + molecular + treatment for complex ML pipelines. 9. **Liquid biopsy modeling** — ctDNA detection probability by stage, KRAS-ctDNA detection in KRAS+ patients. 10. **Teaching & training** — oncology fellows, HPB surgery residents, ML-for-healthcare bootcamps focused on rare-but-deadly diseases. --- ## Loading examples ### pandas (all 5 tables) ```python import pandas as pd df = pd.read_csv("hconc005_sample.csv") ca_long = pd.read_csv("hconc005_ca19_9_longitudinal.csv") surg = pd.read_csv("hconc005_surgical.csv") mol = pd.read_csv("hconc005_molecular.csv") tx = pd.read_csv("hconc005_treatment_history.csv") print(df.shape) # (500, 107) print(ca_long.shape) # (~4,400, 3) print(surg.shape) # (~100, 21) print(df["kras_mutation"].value_counts()) ``` ### Hugging Face `datasets` ```python from datasets import load_dataset ds = load_dataset("xpertsystems/hconc005-sample") df = ds["train"].to_pandas() ``` ### KRAS variant survival analysis ```python from lifelines import KaplanMeierFitter import matplotlib.pyplot as plt df["dead"] = (df["vital_status"] != "Alive").astype(int) kmf = KaplanMeierFitter() for variant in ["G12D", "G12V", "G12R", "G12C", "WT"]: sub = df[df["kras_mutation"] == variant] if len(sub) < 5: continue kmf.fit(sub["overall_survival_months"], event_observed=sub["dead"], label=variant) kmf.plot_survival_function() plt.title("OS by KRAS Variant in PDAC") plt.show() ``` ### CA 19-9 trajectory by RECIST response ```python import matplotlib.pyplot as plt import numpy as np merged = ca_long.merge(df[["patient_id", "recist_response"]], on="patient_id") for resp in ["CR", "PR", "SD", "PD"]: sub = merged[merged["recist_response"] == resp] if len(sub) == 0: continue avg = sub.groupby("month_from_diagnosis")["ca19_9_u_ml"].median() plt.plot(avg.index, avg.values, label=resp, marker='o', markersize=3) plt.yscale("log") plt.xlabel("Months from diagnosis"); plt.ylabel("CA 19-9 (U/mL, log scale)") plt.legend(); plt.title("Median CA 19-9 Trajectory by RECIST Response") plt.show() ``` ### HRR-stratified olaparib uptake audit ```python hrr_pos = df[df["parp_inhibitor_eligible_flag"] == 1] ola_uptake = (hrr_pos["second_line_regimen"] == "Olaparib").mean() print(f"Olaparib in PARP-eligible second-line: {ola_uptake:.1%}") # Expected ~25-40% (POLO era) ``` ### Whipple complications by approach ```python whipple = surg[surg["surgery_procedure"].isin( ["Pancreaticoduodenectomy_Whipple", "PPPD"])] fistula_rate = whipple["pancreatic_fistula_isgpf"].isin( ["Grade_B", "Grade_C"]).mean() print(f"Whipple ISGPF Grade B/C fistula: {fistula_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. **🚨 Generator required a wrapper patch to run at all.** The original generator had a fatal `ValueError` at line 60: ```python patient_ids = [str(uuid.UUID(int=rng.integers(0, 2**128))) for _ in range(N)] ``` Numpy's `rng.integers()` returns int64 values, which max out at ~`2^63`. Passing `2**128` as `high` raises `ValueError: high is out of bounds for int64`. The wrapper applies a minimal in-place patch that composes a 128-bit UUID from four 32-bit integer draws (deterministic-seeded): ```python _uuid_parts = rng.integers(0, 2**32, (N, 4), dtype=np.int64) patient_ids = [str(uuid.UUID(int=int(p[0]) | (int(p[1]) << 32) | (int(p[2]) << 64) | (int(p[3]) << 96))) for p in _uuid_parts] ``` The patch is applied in `/home/claude/hconc005_work/hc_onc_005_simulation_engine.py`; the original is preserved at `hc_onc_005_simulation_engine_original.py.bak`. **The full commercial product fixes this in the source.** 2. **Module-level script architecture.** Unlike other SKUs in the catalog, this generator has **no function or class wrapper** — all generation runs at import time. `parser.parse_args()` executes at module level, reading from `sys.argv`. The wrapper patches `sys.argv` before each `import` / `importlib.reload()` cycle. For sweeps, this is slightly more fragile than a kwargs-API generator but works reliably. 3. **OS median heavily Stage IV-weighted.** Cohort is ~47% Stage IV at dx (matches SEER ~50-55%), so overall median OS ~5.7mo trends toward the PDAC mortality reality. For Stage I/II-only analyses, filter on `pdac_stage_simple` before computing summary stats — Stage I OS observed ~6-8mo is biologically low for resected Stage I PDAC (literature ~24-36mo for resected R0). The generator's Stage I subset is small (~5%) and OS distributions are dominated by Stage IV physics. **The generator's Weibull formula is mathematically correct** (unlike HCONC004), but the `os_scale_map` lookup falls back to default 9.0 months for many Stage I combinations not explicitly mapped. 4. **Stage I cohort is small (~3-5% of 500 patients = 15-25 patients).** Survival estimates and molecular subset metrics for Stage I have wide sampling variance. **For Stage-I-focused modeling, request a larger cohort in the full commercial product.** 5. **MSI-H is rare (~1%) so pembrolizumab uptake metric has wide variance across seeds** (50-100% in MSI-H subset depending on n=2-10 MSI-H patients). The structural floor is set at ≥40% to accommodate this. 6. **`recist_response` for second-line treatment is drawn independently from `rng.choice(["PR","SD","PD"], p=[0.15, 0.35, 0.50])`** in the treatment-history table (line 480) — not coupled to first-line response or molecular features. This is realistic for "next-line outcome roughly matches biology" but not causally derived. 7. **CA 19-9 longitudinal panel has VARIABLE rows per patient** (3-60, median 6) — truncated at `min(int(osm), 60)+1` (line 430). Shorter survivors have fewer CA 19-9 visits. **Cannot use this panel for fixed-N visit analyses without filtering.** Join on `patient_id` and `groupby` is safe. 8. **`time_to_next_treatment_months`** is drawn from an exponential with rate 0.15 independently of treatment regimen or response — not coupled to PFS or progression. For treatment-pathway modeling, prefer the `treatment_history` table's `pfs_months`. 9. **`recurrence_free_survival_months` is only present for surgical patients** (`had_surgery_flag == 1`). For non-surgical patients, this column is `NaN` by design. 10. **Lewis-negative patients (~10%) have low CA 19-9 even with disease** (line 171: `np.where(lneg_fl==1, rng.uniform(0.1, 5.0, N), ...)`). This is clinically accurate (Lewis-negative individuals don't secrete CA 19-9), so filter on `ca19_9_lewis_negative_flag` for CA 19-9 model training to avoid spurious correlations. 11. **No CEA longitudinal panel** — only CA 19-9 has the longitudinal table. CEA is captured as `cea_baseline_ng_ml` only in the primary cohort. 12. **Race/ethnicity not coupled to outcomes.** Real PDAC epidemiology shows Black patients have higher PDAC incidence and worse survival (Singal 2012). Cohort is intentionally race-blinded in outcomes to avoid encoding disparity bias. 13. **`dx_year` uniform 2015-2026.** Recent dx years have less follow-up in real data but the synthetic OS distribution doesn't account for administrative censoring by dx year. For temporal trend analyses, use only `dx_yr <= 2021` patients. 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) | | CA 19-9 panel | ~4,400 rows (variable) | Configurable cadence | | Int64 UUID bug | Wrapper-patched (disclosed) | **FIXED** in source | | Module-level CLI | Yes (script-style) | Function-API option | | Stage I cohort size | ~5% (small subset) | Configurable enrichment | | Race-outcome coupling | None (race-blinded) | Configurable disparity profiles | | CEA longitudinal | Baseline only | Full longitudinal panel option | | Validation report | Yes (35 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-PAAD | | Multi-line treatment | First + Second | Multi-line cascade | | Support | Community | Email / SLA | --- ## Citation ```bibtex @dataset{xpertsystems_hconc005_2026, title = {HC-ONC-005: Pancreatic Cancer (PDAC) Synthetic Cohort with CA 19-9 Longitudinal Panel and Five Relational Tables}, author = {{XpertSystems.ai}}, year = {2026}, version= {1.0.0}, url = {https://huggingface.co/datasets/xpertsystems/hconc005-sample}, license= {CC-BY-NC-4.0 (sample); Commercial (full product)}, note = {Calibrated against SEER PDAC 2017-2021, TCGA PAAD molecular frequencies (Waddell 2015, Bailey 2016), NCCN Pancreatic Adenocarcinoma Guidelines 2024, AJCC 8th Edition, ISGPF 2016 fistula grading, ISGPS DGE grading, PRODIGE-4 / ACCORD-11 (Conroy 2011 FOLFIRINOX), MPACT (Von Hoff 2013 Gem+NabP), POLO (Golan 2019 olaparib maintenance), NAPOLI-3 (Wainberg 2023 NALIRIFOX), KEYNOTE-158 (Marabelle 2020 pembrolizumab MSI-H), KRYSTAL-1 (Skoulidis 2021 sotorasib G12C), Pritchard 2016 (BRCA2 prevalence), Iacobuzio-Donahue 2009 (SMAD4), Yachida 2010 (metastasis patterns), Hidalgo 2010 (head tumor proportion), Cohen 2018 (PDAC ctDNA).} } ``` --- ## Contact - **Email:** [pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai) - **Web:** [https://xpertsystems.ai](https://xpertsystems.ai) - **Vertical:** Healthcare / Oncology - **SKU catalog:** SKU 5 of the Oncology vertical (15 SKUs total across Cardiology + Oncology); ~80 SKUs across 8 verticals XpertSystems.ai — synthetic data, calibrated to real-world registries.