xpertsystems/hlt004-sample

HLT-004 — Synthetic Disease Progression Dataset (Sample Preview)

A free, schema-identical preview of the full HLT-004 commercial product from XpertSystems.ai.

A fully synthetic longitudinal disease progression dataset combining patient-level baseline records, visit-level biomarker trajectories, and structured event logs across two clinically distinct disease modules: NSCLC (oncology survival) and Heart Failure (chronic cardiovascular progression). Calibrated to SEER 5-year survival, AJCC TNM 8th Edition staging, NYHA Functional Classification, RECIST 1.1 response criteria, CTCAE v5.0 adverse events, and Fine-Gray competing-risk survival outcomes.

⚠️ PRIVACY & SYNTHETIC NATURE Every record in this dataset is 100% synthetic. No real patient data, no PHI, no re-identifiable records. Stage-specific survival distributions and treatment response rates match published SEER / clinical trial benchmarks but the patients are computationally generated.

---

What's in this sample

Two disease modules, each with three CSVs:

nsclc/ — Non-Small Cell Lung Cancer (TNM 8th Edition staging)

File	Rows	Cols	Description
hlt004_baseline.csv	400	37	Patient-level: demographics, ECOG PS, CCI, stage at dx, OS/PFS, death cause, 1L treatment arm, best overall response (CR/PR/SD/PD)
hlt004_longitudinal.csv	3,323	33	Visit-level: ~8.3 visits/patient × 5yr follow-up; CEA biomarker trajectory, labs, AEs, on-treatment flag, RECIST per imaging visit
hlt004_events.csv	1,943	4	Event log: diagnosis / treatment_start_1L / treatment_end_1L / response_assessment / progression / os_endpoint

heart_failure/ — Heart Failure (NYHA Class I-IV)

File	Rows	Cols	Description
hlt004_baseline.csv	400	37	Patient-level: demographics, ECOG-equivalent ambulation, CCI, NYHA Class at dx, OS/PFS-equivalent, death cause, 1L therapy arm (GDMT/Ivabradine/LVAD)
hlt004_longitudinal.csv	4,502	33	Visit-level: ~11.3 visits/patient over 5yr; NT-proBNP biomarker, labs, AEs
hlt004_events.csv	1,969	4	Event log analogous to NSCLC structure

Total: ~12,500 rows across 8 CSVs + 2 generator summaries = ~2.2 MB.

---

Schema highlights

Baseline (37 columns, patient-level)

Identity: patient_id, disease_type, diagnosis_date

Demographics: age_at_dx, sex, race_ethnicity, bmi_baseline, smoking_ever, pack_years, insurance_type

Clinical scoring: ecog_ps_baseline (0-4), cci_at_dx (Charlson Comorbidity Index), stage_at_dx, stage_idx_at_dx (numeric)

Survival (Fine-Gray competing risk): os_days, os_event (death/censored), pfs_days, pfs_event, censoring_reason, death_cause, last_contact_days

Treatment: treatment_arm_1L (regimen code), best_overall_response (CR/PR/SD/PD per RECIST 1.1), response_day_from_dx, progression_day_from_dx

Longitudinal (33 columns, visit-level)

Visit metadata: patient_id, visit_id, visit_number, visit_date, days_from_dx, visit_type (baseline/imaging/routine_lab/urgent/hospitalization)

Disease state: stage_current, stage_idx_current, stage_changed_flag (Markov transitions)

Biomarker (disease-specific): NSCLC → cea_value (CEA, ng/mL); HF → ntprobnp_value (NT-proBNP, pg/mL); plus *_units

Treatment state: on_treatment_flag, treatment_line, treatment_arm

RECIST 1.1: recist_response_visit (CR/PR/SD/PD/N/A per imaging visit, derived from marker change)

Labs (10 panels): wbc, hgb, plt, anc, creatinine, bun, alt, ast, albumin, bilirubin

Adverse events (CTCAE v5.0): ae_count_grade1, ae_count_grade2, ae_count_grade3, ae_count_grade4

Events (4 columns, sparse event log)

patient_id, event_type ∈ {diagnosis, treatment_start_1L, treatment_end_1L, response_assessment, progression, os_endpoint}, event_day (days from dx), event_detail

---

Coverage

2 fully-calibrated disease modules demonstrated:

• NSCLC (oncology) — TNM 8th Edition, SEER-calibrated OS by stage, CEA biomarker, 4 treatment regimens (Carboplatin/Pemetrexed/Pembrolizumab, Carboplatin/Paclitaxel, Pembrolizumab mono, Docetaxel 2L)
• Heart Failure (chronic CV) — NYHA Class I-IV, 3 treatment lines (GDMT, Ivabradine add-on, LVAD)

Survival modeling:

• Weibull-distributed time-to-event with shape parameter 1.2
• Stage-specific median OS calibrated to SEER (NSCLC) and Pocock et al. 2013 (HF)
• Competing-risk death cause assignment (Fine-Gray framework)
• Administrative censoring at follow-up end

Biomarker dynamics:

• Gaussian-process trajectories with stage-mean drift
• Response-driven dips (CR/PR) and progression-driven rises
• Disease-specific markers calibrated to clinical literature

Stage transitions:

• Markov-chain time-inhomogeneous transitions
• Stage 1→2→3→4 progression probabilities tied to PFS days

Adverse events:

• CTCAE v5.0 grade 1-4 counts per visit
• On-treatment elevation, ECOG-driven baseline rates

Missing data:

• MAR rate 10% + MCAR rate 2% applied to biomarker/lab columns

---

Calibration source story

The full HLT-004 generator anchors all distributions to authoritative oncology and cardiology references:

• SEER (Surveillance, Epidemiology, End Results) Program — 5-year overall survival benchmarks by cancer stage (NSCLC Stage I=63%, II=37%, III=15%, IV=6%)
• AJCC TNM 8th Edition (2017) — Cancer staging anatomical definitions
• NYHA Functional Classification (1994 update) — Heart Failure stages I-IV with associated 1-year mortality 5%/15%/30%/50%
• Roche et al. (2020) Lancet — Advanced NSCLC median OS 10-14 months with platinum chemotherapy
• Pocock et al. (2013) — MAGGIC heart failure risk model; chronic HF mortality predictors
• RECIST 1.1 (Eisenhauer et al. 2009) — Response Evaluation Criteria In Solid Tumors
• CTCAE v5.0 (NCI) — Common Terminology Criteria for Adverse Events
• Fine & Gray (1999) — Competing risk regression for cause-specific mortality
• CONSORT 2010 — Longitudinal data hygiene conventions

Sample-scale validation scorecard

Metric	Observed	Target	Tolerance	Status	Source
NSCLC overall mortality rate	76.5%	77%	±10%	✅ PASS	Roche et al. (2020)
NSCLC median OS (months)	13.8	12.0	±4.0	✅ PASS	SEER 2021
HF overall mortality rate	62.5%	63%	±10%	✅ PASS	Pocock et al. (2013)
HF median OS (months)	19.5	18.0	±6.0	✅ PASS	Pocock et al. (2013)
RECIST response categories	4	4	—	✅ PASS	RECIST 1.1
TNM stage diversity	4	4	—	✅ PASS	AJCC TNM 8th
Longitudinal temporal monotonicity	100%	100%	±2%	✅ PASS	CONSORT
PFS ≤ OS invariant	100%	100%	±2%	✅ PASS	Fine & Gray (1999)
Event log completeness	100%	100%	±5%	✅ PASS	Data hygiene
Missing data rate	11.8%	12%	±6%	✅ PASS	MAR 10% + MCAR 2%

Grade: A+ (100/100) — verified across 6 random seeds (42, 7, 123, 2024, 99, 1).

---

Loading examples

Pandas

import pandas as pd

base = pd.read_csv("nsclc/hlt004_baseline.csv")
long = pd.read_csv("nsclc/hlt004_longitudinal.csv")
events = pd.read_csv("nsclc/hlt004_events.csv")

print(f"Patients: {len(base)}")
print(f"Visits: {len(long)}")
print(f"Events: {len(events)}")

# Survival by stage
print("\nNSCLC OS by stage (event rate, median OS days):")
for stage, grp in base.groupby("stage_at_dx"):
    events_only = grp[grp["os_event"] == 1]
    if len(events_only):
        print(f"  {stage}: n={len(grp)}  "
              f"event_rate={grp['os_event'].mean():.2%}  "
              f"median_os={events_only['os_days'].median():.0f}d")

Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset("xpertsystems/hlt004-sample", data_files={
    "nsclc_baseline":      "nsclc/hlt004_baseline.csv",
    "nsclc_longitudinal":  "nsclc/hlt004_longitudinal.csv",
    "nsclc_events":        "nsclc/hlt004_events.csv",
    "hf_baseline":         "heart_failure/hlt004_baseline.csv",
    "hf_longitudinal":     "heart_failure/hlt004_longitudinal.csv",
    "hf_events":           "heart_failure/hlt004_events.csv",
})
print(ds)

Kaplan-Meier survival curve

import pandas as pd
import matplotlib.pyplot as plt
# Optional: pip install lifelines
from lifelines import KaplanMeierFitter

base = pd.read_csv("nsclc/hlt004_baseline.csv")
fig, ax = plt.subplots(figsize=(8, 5))
for stage, grp in base.groupby("stage_at_dx"):
    kmf = KaplanMeierFitter()
    kmf.fit(grp["os_days"], event_observed=grp["os_event"], label=stage)
    kmf.plot_survival_function(ax=ax)
ax.set_xlabel("Days from diagnosis"); ax.set_ylabel("Overall Survival")
ax.set_title("NSCLC OS by TNM Stage")
plt.show()

Time-varying covariate Cox model

import pandas as pd
from lifelines import CoxTimeVaryingFitter

long = pd.read_csv("nsclc/hlt004_longitudinal.csv")
base = pd.read_csv("nsclc/hlt004_baseline.csv")[["patient_id", "os_days", "os_event"]]

# Build start/stop format for time-varying biomarker
long_sorted = long.sort_values(["patient_id", "days_from_dx"])
long_sorted["start"] = long_sorted.groupby("patient_id")["days_from_dx"].shift(0)
long_sorted["stop"] = long_sorted.groupby("patient_id")["days_from_dx"].shift(-1)
long_sorted = long_sorted.merge(base, on="patient_id")
long_sorted["event_at_stop"] = (long_sorted["stop"] >= long_sorted["os_days"]) & (long_sorted["os_event"] == 1)
long_sorted = long_sorted.dropna(subset=["stop", "cea_value"])

ctv = CoxTimeVaryingFitter()
ctv.fit(long_sorted[["patient_id", "start", "stop", "event_at_stop", "cea_value"]]
        .rename(columns={"event_at_stop": "event"}),
        id_col="patient_id", event_col="event", start_col="start", stop_col="stop")
ctv.print_summary()

---

Suggested use cases

• Survival modeling — Cox proportional hazards, Kaplan-Meier, accelerated failure time, parametric survival (Weibull, log-normal, log-logistic)
• Time-varying covariate analysis — joint longitudinal-survival models using biomarker trajectories
• Competing-risk regression — Fine-Gray subdistribution hazard models on death_cause field
• Stage transition modeling — multi-state Markov models on stage_current evolution
• Biomarker trajectory clustering — latent class growth analysis on CEA / NT-proBNP series
• Treatment effect estimation — propensity score / inverse probability weighting from observational treatment_arm_1L assignment
• RECIST response prediction — predict CR/PR/SD/PD from baseline + early biomarker dynamics
• Adverse event hazard modeling — predict Grade 3+ AE onset from on-treatment patient state
• Missing data methodology — develop MAR/MCAR/MNAR imputation strategies on flagged data
• ML model pretraining — pretrain healthcare survival models before fine-tuning on real registry data (SEER, CIBMTR, etc.)

---

Sample vs. full product

Aspect	This sample	Full HLT-004 product
Disease modules	2 (NSCLC + HF)	4 fully-calibrated + 11 templated (15 total)
Patients per disease	400	10,000+ (default) up to 100K
Follow-up window	5 years	Configurable 1-15 years
Schema	identical	identical
Calibration	identical	identical
License	CC-BY-NC-4.0	Commercial license

The full product includes:

• All 4 fully-calibrated diseases: NSCLC, Heart Failure, CKD, and Breast Cancer (after stage-distribution fix)
• 11 additional disease modules (Colorectal, Prostate, Ovarian, COPD, T2DM Progressive, Multiple Sclerosis, Alzheimer's, Hepatocellular Carcinoma, AML, Rheumatoid Arthritis, HIV/AIDS) — these currently use NSCLC defaults as templates pending per-disease calibration
• Up to 100K patients per disease for production-grade model training
• Configurable follow-up windows up to 15 years

Contact us for the full product and calibration roadmap.

---

Limitations & honest disclosures

• Sample is preview-only. 400 patients per disease × 5yr follow-up is enough to demonstrate schema, calibration, and survival shape, but is not statistically sufficient for serious prognostic model training, especially for rare death-cause analysis or biomarker discovery. Use the full product (10K+ patients per disease) for serious work.
• Two disease modules in this sample, not all 15. The full HLT-004 catalogue lists 15 disease keys, but currently only 4 (NSCLC, Heart Failure, CKD, Breast Cancer) have per-disease calibration; the remaining 11 use NSCLC defaults as placeholder templates. This sample includes the 2 most-requested modules (NSCLC for oncology, Heart Failure for chronic CV).
• breast_cancer module has a known stage-distribution normalization bug in the current generator release (stage_dist_dx sums to 0.94 instead of 1.0, causing a NumPy ValueError when sampling). This is a single one-line fix in the generator's catalogue but is not yet patched in the version distributed with this sample. We excluded breast_cancer from this preview rather than patch around it. Will be addressed in the next generator release.
• Visit-level recist_response_visit is change-detection, not best-response. RECIST 1.1 best response per patient is in baseline.csv → best_overall_response. The visit-level field compares consecutive biomarker readings to flag changes during follow-up; it rarely yields "CR" since CR requires lesion disappearance (not capturable from a single marker comparison). Use best_overall_response for cohort-level response analysis.
• Biomarker trajectories are simulation, not real labs. Stage-specific drift and response/progression-driven trajectories follow published clinical patterns (CEA elevation with NSCLC progression, NT-proBNP elevation with HF decompensation) but do NOT capture the full noise structure, batch effects, or measurement variability of real lab data. Use for ML algorithm development; validate on real registry data.
• Treatment assignment is propensity-weighted, not randomized. Treatment arms are assigned based on stage, ECOG PS, CCI, and insurance — a non-randomized mechanism reflecting real-world prescribing. Causal effect estimation requires propensity score adjustment, IPW, or G-methods.
• Stage transitions are Markov, not informative. Stage changes follow a memoryless transition kernel; the full real-world progression dynamics include informative censoring patterns not captured here.
• Death causes are simplified. death_cause field includes the primary cause; real death certificates have ICD-10 underlying + contributing causes which are not in this sample.
• CCI (Charlson) score is baseline only. Real comorbidity index evolves over follow-up; the sample treats it as fixed at diagnosis.

---

Ethical use guidance

This dataset is designed for:

• Survival analysis methodology development
• Healthcare ML model pretraining
• Educational use in oncology biostatistics and HF outcomes research
• Synthetic data validation methodology research
• ETL pipeline testing for clinical registries (SEER-conformant schemas)

This dataset is not appropriate for:

• Making decisions about real individual patients
• Clinical prognosis claims of any kind
• Drug efficacy or treatment comparisons in regulatory submissions
• Training models that produce real clinical recommendations
• Discriminatory analyses on protected demographic groups

---

Companion datasets in the Healthcare vertical

• HLT-001 — Synthetic Patient Population (5K patients × 79 cols, CDC/NHANES calibrated)
• HLT-002 — Synthetic EHR Dataset (4K encounters + FHIR R4 bundles)
• HLT-003 — Synthetic Clinical Trial Dataset (3 endpoint types + power sweep)
• HLT-004 — Synthetic Disease Progression Dataset (you are here)

Use HLT-001 → HLT-004 together for population → encounter → trial → longitudinal-progression healthcare ML workflows.

---

Citation

If you use this dataset, please cite:

@dataset{xpertsystems_hlt004_sample_2026,
  author       = {XpertSystems.ai},
  title        = {HLT-004 Synthetic Disease Progression Dataset (Sample Preview)},
  year         = 2026,
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/xpertsystems/hlt004-sample}
}

---

Contact

• Web: https://xpertsystems.ai
• Email: pradeep@xpertsystems.ai
• Full product catalog: Cybersecurity, Insurance & Risk, Materials & Energy, Oil & Gas, Healthcare, and more

Sample License: CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0) Full product License: Commercial — please contact for pricing.