Synthetic Oncology Clinical Trial Dataset

Synthetic CDISC-compliant datasets modeled after NCT01797120 (PrE0102) — a Phase II randomized trial of Fulvestrant ± Everolimus in postmenopausal HR+ metastatic breast cancer.

All data is purely synthetic. Study ID NCT01797120. No real patient data is present in this repository.

---

Source Trial Summary

Field	Value
Trial	NCT01797120 (PrE0102)
Indication	HR+ HER2− metastatic breast cancer (AI-resistant, postmenopausal)
Design	Phase II, randomized, double-blind, placebo-controlled
Arms	Fulvestrant 500 mg + Everolimus 10 mg (n=100) vs. Fulvestrant + Placebo (n=100)
Primary endpoint	Progression-Free Survival (PFS)
Published PFS	10.3 months (treatment) vs. 5.1 months (placebo)
Publication	Journal of Clinical Oncology, June 2018 (PMID: 29664714)

---

Repository Structure

bc-supporting-synthetic-data/
├── .github/
│   └── workflows/
│       └── lint.yml                    # GitHub Actions: black + flake8 on every push/PR
├── src/
│   └── cdisc_generation_functions.py   # Python functions that generate all datasets
├── datasets/                           # Generated CDISC CSV files (see Datasets section)
├── skills/
│   └── validate-synth-data/            # Claude Code skill for dataset validation
│       ├── SKILL.md                    # Skill definition and workflow
│       ├── scripts/validate.py         # 100-check validation script
│       ├── evals/evals.json            # Skill test prompts
│       └── references/                 # CDISC rule sets and RECIST 1.1 supplement
├── protocol/
│   └── NCT01797120.pdf                 # Original trial protocol
├── results/
│   ├── NCT01797120-results.md          # Human-readable results summary
│   └── NCT01797120-results.fhir.json   # FHIR EBM bundle of published results
├── .pre-commit-config.yaml             # Pre-commit hooks (black, flake8)
├── ONBOARDING.md                       # Onboarding guide for new contributors
├── requirements.txt                    # Python dependencies
├── setup.cfg                           # Flake8 and black configuration (max-line-length=120)
└── validation_report.md                # Latest validation output (100/100 checks)

---

Datasets

SDTM Domains

Domain	File	Description
DM	DM.csv	Demographics — age, sex, arm, randomization/consent/reference dates
EX	EX.csv	Exposure — drug, dose, route, start/end dates, study days
TU	TU.csv	Tumor Identification — target lesion locations per RECIST 1.1 (TU domain)
TR	TR.csv	Tumor Results — per-lesion LDIAM and visit-level SUMDIAM measurements
RS	RS.csv	Disease Response — CR/PR/SD/PD/NE derived from SUMDIAM per RECIST 1.1
DS	DS.csv	Disposition — informed consent, randomization, treatment stop, study end
AE	AE.csv	Adverse Events — fatal SAE (AESER, AESDTH, AEOUT="FATAL") for one Placebo subject's death
FA	FA.csv	Findings About — occurrence indicators for DS disposition/milestone events (FAOBJ = DSDECOD)
RELREC	RELREC.csv	Related Records — links FA and DS records per subject
TV	TV.csv	Trial Visits — planned visit schedule through ~18 cycles
TA	TA.csv	Trial Arms — SCREENING / INDUCTION / CONTINUATION / FOLLOW-UP epochs

ADaM Datasets

Dataset	File	Description
ADSL	ADSL.csv	Subject-Level Analysis — baseline, flags (ITTFL, SAFFL, PPROTFL), PFS, best response
ADTTE	ADTTE.csv	Time-to-Event — PFS with AVAL (days), CNSR, PARAMCD="PFS"

Key Variables

Variable	Description
USUBJID	Unique subject ID — format NCT01797120-NNNN
TRT01A	Actual treatment arm — "Treatment" or "Placebo"
RSSTRESC	RECIST 1.1 response — CR, PR, SD, PD, or NE
AVAL	Analysis value (days to progression or censoring) in ADTTE
CNSR	Censoring flag — 0 = progression event, 1 = censored
TRLOBXFL	Last observation before exposure flag (baseline) in TR
TRLNKID	Links TR lesion records to TU tumor identification records

---

Regenerating the Data

Prerequisites

pip install -r requirements.txt
# or, using the project venv:
python -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt

Run the generator

python src/cdisc_generation_functions.py

This writes all CSV files to datasets/ using numpy.random.seed(0) for reproducibility.

Generation order (if calling functions directly)

Each function depends on the previous output:

from src.cdisc_generation_functions import *
import numpy as np

np.random.seed(0)

dm     = create_dm(n=200)
ex     = create_ex(dm)
events = derive_events(ex, dm)
ex     = finalize_ex(ex, events, dm)   # cap EX end dates at progression
tr     = create_tr(ex, events, dm)
rs     = derive_rs(tr)
tu     = create_tu(dm, tr)
dm     = finalize_dm(dm, ex, events)
ds     = create_ds(dm, ex, events)
adsl   = create_adsl(dm, ex, rs, events)
adtte  = create_adtte(adsl)
tv     = create_tv()
ta     = create_ta()

---

Validating the Datasets

A dedicated validation script checks 100 rules across seven categories:

Category	Checks
CDISC Structural	Required variables, controlled terminology, value formats per domain
Cross-domain Integrity	USUBJID consistency, arm-drug assignments, DS record completeness
RECIST Derivation	SUMDIAM traceability, response classification, TU-TR linkage
Timeline & Death Logic	Date ordering, no post-death observations, DTHDTC consistency
Population Distribution	Age range, enrollment spread, variability across subjects
Duplicate & Study Consistency	Sequence key uniqueness, domain coverage
PFS Fidelity	Observed median PFS vs. published trial results (±20% tolerance)

Run directly

python skills/validate-synth-data/scripts/validate.py \
  --datasets datasets/ \
  --output validation_report.md

Run via Claude Code

If you have the skill installed (see Claude Code Setup below), you can ask Claude to validate the datasets in plain language:

"validate my datasets"
"check the CSVs before I package the ZIP"
"something looks off with the RS domain — audit the RECIST derivation"

Claude will run the script and present the report with plain-language explanations of any failures.

---

Claude Code Setup

This project includes a Claude Code skill (skills/validate-synth-data/) that adds dataset validation as a natural-language command. To activate it:

1. Install Claude Code

npm install -g @anthropic-ai/claude-code

Or download the desktop app from claude.ai/code.

2. Install the skill

Copy (or symlink) the skill into your Claude Code skills directory:

# Copy
cp -r skills/validate-synth-data ~/.claude/skills/validate-synth-data

# Or symlink (changes in the repo are reflected immediately)
ln -s "$(pwd)/skills/validate-synth-data" ~/.claude/skills/validate-synth-data

3. Open the project in Claude Code

claude  # from the project root

In the Claude session, run:

/init

This will prompt you for instructions and create the CLAUDE.md file.

Claude will pick up the skill from ~/.claude/skills/.

4. Verify the skill is active

In the Claude Code session, run:

"validate my datasets"

You should see a report confirming 100/100 checks pass against the committed datasets.

Updating the skill

If you modify skills/validate-synth-data/scripts/validate.py and used a symlink, the changes are live immediately. If you copied, re-run the cp command above.

---

CDISC Standards Conformance

The datasets are generated to conform to the following standards. The validation_report.md in this repository reflects the current conformance status.

Standard	Version	Coverage
SDTMIG	3.4	DM, EX, TU, TR, RS, DS, FA, RELREC, TV, TA
ADaMIG	1.3	ADSL (OCCDS), ADTTE (TTE)
RECIST	1.1	LDIAM/SUMDIAM in TR; CR/PR/SD/PD/NE in RS; TU domain
CDISC Conformance Rules	CORE	Key rules: CG0102, CG0073, CG0066, CG0432, CG0142, FB2201, FB0611, CG0563

---

License

Synthetic data only. No patient privacy concerns. Refer to protocol/NCT01797120.pdf for the original trial protocol (copyright-exempt for non-commercial use per the RECIST Working Group).