README.md

---
license: cc-by-nc-4.0
language:
- en
tags:
- healthcare
- endocrinology
- diabetes
- type-2-diabetes
- t2dm
- synthetic-data
- pharmacotherapy
- glp1
- sglt2
- ehr
- clinical
- fhir
pretty_name: "HC-END-002 Type 2 Diabetes Synthetic Dataset (Sample)"
size_categories:
- n<1K
task_categories:
- tabular-classification
- tabular-regression
---

# HC-END-002 — Type 2 Diabetes Synthetic Dataset (Sample)

**XpertSystems.ai · Synthetic Data Factory · Endocrinology Vertical**

A wide, physiologically grounded synthetic cohort of Type 2 Diabetes (T2DM) patient
records spanning demographics, anthropometrics, a full metabolic/lipid/hepatic panel,
blood pressure, multi-class pharmacotherapy (metformin, SGLT2i, GLP-1 RA, DPP-4i,
sulfonylureas, TZDs, insulin), lifestyle, disease progression, micro/macrovascular
complications, NAFLD/NASH, and self-management. This repository contains a **500-row,
single-seed sample**. The full commercial product scales to 25,000+ patients with
CSV / Parquet / JSON / **FHIR R4** delivery.

- **SKU:** HC-END-002
- **Sample size:** 500 patients × 140 columns
- **License (sample):** CC-BY-NC-4.0 — commercial license available for the full product
- **Contact:** pradeep@xpertsystems.ai · https://xpertsystems.ai

---

## Validation

This sample passes XpertSystems Grade **A+** validation (overall **10.000 / 10**) with
deterministic reproduction across all six canonical seeds `[42, 7, 123, 2024, 99, 1]`.

Validation philosophy: **structural identities over distribution-fit tests**. Scorecard
ranges are anchored to named T2DM sources and calibrated to *observed* engine behavior.
Where the engine diverges from population literature, the divergence is **disclosed
below** rather than hidden — the sweep stays deterministic without masking the gaps.

### Calibration anchors

| Metric | Sample value | Target range | Source |
|---|---|---|---|
| Mean BMI | 32.9 kg/m² | 31–35 | NHANES 2017-2020 T2DM (~32-34) |
| Mean diabetes duration | 12.7 yr | 10–15 | Engine exponential duration model |
| MACE prevalence | 19.6% | 12–26% | UKPDS / ACCORD / EMPA-REG / LEADER |
| Metformin use | 73.6% | 68–85% | ADA Standards of Care 2025 (first-line ~78%) |
| Insulin use | 37.6% | 30–50% | Duration-driven (~25-50% at 10yr) |
| GLP-1 RA adoption | 30.4% | 20–36% | LEADER-era uptake (~28%) |
| SGLT2i adoption | 34.8% | 28–42% | EMPA-REG-era uptake (~35%) |
| CKD prevalence (eGFR<60) | 40.8% | 34–48% | Long-duration cohort |
| Neuropathy prevalence | 33.8% | 26–42% | DSPN ~30-40% long-standing T2DM |
| Retinopathy prevalence | 45.8% | 38–52% | Engine observed |
| **HOMA-IR > 2 (insulin resistance)** | **100%** | **≥0.95** | Near-universal in T2DM |
| **eGFR in [15,125]** | **100%** | **≥1.0** | Renal physiology bounds |
| **LDL > 0 (Friedewald)** | **100%** | **≥1.0** | Lipid integrity |
| **Retinopathy stage well-formed** | **100%** | **≥1.0** | Schema integrity |
| **Column count** | **140** | **≥130** | Schema completeness |

---

## Schema highlights by module (140 columns)

**Demographics & anthropometrics.** Sex, age at diagnosis, duration, current age,
race/ethnicity, insurance; BMI, weight, height, waist circumference, waist-hip ratio,
body-fat %, visceral-fat index.

**Metabolic panel.** Baseline & current HbA1c, fasting/postprandial glucose, fasting
insulin, HOMA-IR, HOMA-B, C-peptide, beta-cell function %; full lipids (TC/LDL/HDL/TG/
non-HDL); uric acid, CRP, ALT/AST/GGT, adiponectin, leptin, TSH.

**Blood pressure.** SBP/DBP, hypertension flag, antihypertensive class.

**Pharmacotherapy (multi-class).** Pharmacotherapy line; metformin (dose, A1C reduction);
SGLT2i (drug, A1C/weight/SBP effects, UTI); GLP-1 RA (drug, A1C/weight effects, nausea/
vomiting); DPP-4i; sulfonylurea (hypoglycemia events, weight gain); TZD; insulin (type,
basal/total units, A1C reduction, weight gain); adherence %, polypharmacy count.

**Lifestyle.** Diet pattern & quality, caloric intake, carb %, fiber, saturated fat;
exercise type/frequency/minutes, sedentary hours, VO₂max; sleep hours/quality, AHI/OSA;
smoking, pack-years, alcohol, stress, weight-loss intervention & 1-yr weight change.

**Disease progression & complications.** Markov diabetes stage; retinopathy (stage,
time-to-event); nephropathy (eGFR, UACR, micro/macroalbuminuria, CKD, ESRD); neuropathy
(type, DNSS); cardiovascular (MACE, CAD, stroke, HF, PAD); NAFLD/NASH/cirrhosis;
dyslipidemia & statin therapy.

**Self-management.** BG monitoring frequency, CGM use/device, diabetes education, endo
referral, visit counts, foot/eye exams, distress, self-efficacy, health literacy, social
support, financial toxicity.

**Coding standards.** ICD/SNOMED (44054006, DM type 2), LOINC (4548-4, HbA1c); ships a
FHIR R4 Bundle (Patient / Condition / Observation) in the full product.

---

## Files

- `hc_end_002_sample.csv` — 500-patient sample (140 columns)
- `generate_sample_dataset_hc_end_002.py` — reproducible generator + validation harness
- `validation_report.json` / `validation_report.md` — full scorecard
- `sweep_summary.json` — 6-seed determinism results

## Loading

```python
import pandas as pd
df = pd.read_csv("hc_end_002_sample.csv")
print(df[["patient_id","hba1c_current_pct","bmi_kg_m2",
          "pharmacotherapy_line","glp1_drug","mace_flag"]].head())
```

```python
from datasets import load_dataset
ds = load_dataset("csv", data_files="hc_end_002_sample.csv")
```

## Use cases

- Pharmacotherapy-pathway and treatment-effect modeling across drug classes
- CV-risk and complication-onset prediction (UKPDS/ACCORD-style)
- Health-economics and formulary modeling (GLP-1 / SGLT2i uptake scenarios)
- NAFLD/NASH screening-cohort prototyping
- ML training where real T2DM EHR data is PHI-restricted
- FHIR-pipeline development against synthetic R4 bundles (full product)

---

## Honest limitations & disclosed generator behavior

Transparency is a core XpertSystems principle. The v1.0 engine has the following known
deviations from population literature. They are reproducible and disclosed; the data
remains internally consistent for modeling, but users should calibrate expectations.

1. **HbA1c runs low (treatment-effect over-correction).** Per-class A1C reductions
   (metformin + SGLT2i + GLP-1 + DPP-4i + insulin) are summed additively before adherence
   scaling, so net current HbA1c centers near **6.5%** with ~68% at target — versus NHANES
   T2DM means of ~7.5–8% and ~25–35% at target. Treat `hba1c_current_pct` as an optimistic
   "fully-treated" scenario, not a population mean. `hba1c_baseline_pct` (mean ~8.4%) is
   the better population-level anchor.
2. **Hypertension prevalence ~95%.** The flag fires on `SBP≥130 OR DBP≥80` (2017 ACC/AHA
   threshold) against an SBP distribution centered ~135, inflating prevalence above the
   realistic T2DM range (~70–80%). Use the raw SBP/DBP fields if you need a stricter cutoff.
3. **NAFLD prevalence ~79%.** Slightly above the commonly cited ~55–70% for T2DM due to the
   `(rand<0.65 OR BMI>30)` construction. Defensible but on the high side.
4. **Retinopathy stage vs time-to-event mismatch.** ~53% of patients have a retinopathy
   time-to-event ≤ duration ("present"), but only ~46% carry a non-None stage: the stage
   integer is scaled by `duration/20` and floored, demoting some recent-onset present cases
   back to "None". Stage values are well-formed; prevalence is internally ~46%.
5. **Duplicate dict keys.** `beta_cell_function_pct` and an `homa_ir`/`insulin_resistance_index`
   pair are written twice during record assembly; Python keeps the last write, so the columns
   are consistent but the redundancy is noted.

General synthetic-data caveats apply: cross-field correlations beyond those explicitly modeled
may be weaker than in real cohorts. **Not for clinical decision-making** — research/development
use only.

---

## Commercial product comparison

| Capability | This sample | Full HC-END-002 product |
|---|---|---|
| Patients | 500 | 25,000+ (configurable) |
| Seeds / cohorts | 1 | Multi-seed, reproducible |
| Formats | CSV | CSV + Parquet + JSON + **FHIR R4 Bundle** |
| HbA1c calibration | Disclosed (optimistic) | Recalibrated to NHANES population means |
| HTN / NAFLD thresholds | Disclosed (high) | Tunable prevalence targets |
| Retinopathy staging | Approximate | Registry-anchored stage distribution |
| License | CC-BY-NC-4.0 | Commercial |
| Support & SLA | — | Included |

Full product, custom cohorts, or other endocrinology SKUs: **pradeep@xpertsystems.ai**

---

## Citation

```bibtex
@dataset{xpertsystems_hc_end_002_2026,
  title        = {HC-END-002: Type 2 Diabetes Synthetic Dataset},
  author       = {XpertSystems.ai},
  year         = {2026},
  publisher    = {XpertSystems.ai Synthetic Data Factory},
  url          = {https://xpertsystems.ai},
  note         = {Synthetic; CC-BY-NC-4.0 (sample). Calibrated to: NHANES 2017-2020
                  T2DM cohort; UKPDS (UK Prospective Diabetes Study); EMPA-REG OUTCOME
                  (Zinman et al. 2015, NEJM); LEADER (Marso et al. 2016, NEJM);
                  ACCORD (2008) and ADVANCE (2008) glycemic-control trials; ADA
                  Standards of Care in Diabetes 2025; CDC National Diabetes Statistics
                  Report 2024.}
}
```

*Synthetic data generated by XpertSystems.ai. Not derived from real patient records.
Not for clinical use.*