Synthebo

Open source · Apache-2.0 code · CC-BY-4.0 data

Clinical trial data where the right answer is known.

Synthebo generates complete synthetic clinical trials for each therapeutic area, from CDASH raw data through SDTM, ADaM, TLFs and Define-XML, and ships each one with an answer key: the true effect, the correct values and every seeded defect.

Prototype running for three therapeutic areas.

Coverage of 20 simulated trialsTwenty 95% confidence intervals for the treatment effect, plotted against the true effect of -0.60. Nineteen contain it; one misses.-0.9-0.6-0.30.0Estimated treatment effect (HbA1c, %)
  • True effect, set by you
  • 95% CI contains it: 19 of 20
  • Misses, as 1 in 20 should
Illustration: 20 simulated trials of the same design, analyzed with ANCOVA.

Every team needs realistic trial data. Almost none can share it.

Statistical programmers, standards teams and tool builders all need complete studies to build, test and train on. The data that exists doesn't fit the job.

Real patient data can't leave the building

Privacy and confidentiality keep real studies locked inside each sponsor, so they can't be used for open tools, training or public benchmarks.

Public examples are scarce

The best-known public CDISC study is a single legacy pilot in one therapeutic area. There is no open, end-to-end example for oncology, CNS or metabolic studies.

Nobody knows the right answer

With real or de-identified data, the true treatment effect is unknown, so you can't tell whether a program, a pipeline or an AI tool got it right.

Every company rebuilds test data privately

Teams hand-craft dummy datasets for each project, usually for one stage only, and they rarely connect from raw data through to define.xml.

Synthebo's answer: simulate the trial forward from known parameters, so every stage is consistent, shareable and scored against a truth you set yourself.

What Synthebo does

Pick a therapeutic area, set the design, and Synthebo builds the whole study the way it would really be produced, one standard at a time.

  1. Choose a therapeutic area

    Each area comes with its own endpoint, visit schedule and analysis method.

  2. Set the design

    Sample size, allocation, true effect size, dropout and adverse event rates.

  3. Simulate the patients

    Synthebo generates subjects, visits, assessments and events from those parameters.

  4. Generate every stage

    CDASH, SDTM, ADaM, TLFs and Define-XML, plus the answer key, in one download.

What's in every study

One download holds every stage, generated forward from the same simulated patients, so each layer can be checked against the one before it.

  • raw

    CDASH

    Collected data as it would arrive from the sites.

  • tabulation

    SDTM

    Mapped domains with an issue log.

  • analysis

    ADaM

    Derived analysis datasets.

  • outputs

    TLFs

    Tables, listings and figures, including Kaplan-Meier and mean-change plots.

  • metadata

    Define-XML

    Dataset, variable and codelist metadata describing the SDTM and ADaM datasets.

  • ground truth

    Answer key

    The true effect, the correct values and every seeded defect, so results can be scored, not just eyeballed.

Three therapeutic areas today

Each area simulates its own endpoint and analysis. Coverage of the true effect has been checked across repeated seeds.

Oncology

Endpoint
PFS per RECIST
Analysis
Cox hazard ratio
Figure
Kaplan-Meier

Sleep / CNS

Endpoint
Epworth Sleepiness Scale
Analysis
ANCOVA
Figure
Mean change

Metabolic

Endpoint
HbA1c change
Analysis
ANCOVA
Figure
Mean change

Designs planned next:

  • Crossover
  • Randomized withdrawal
  • Single-arm phase 2
  • Group sequential with interims
  • Non-inferiority
  • Open-label extension
  • ISS/ISE pooling

Clean or seeded with errors

Generate a textbook study, or one with known defects planted for your pipeline to find.

Clean

Conformant data at every stage. Use it to test that a pipeline reproduces the true effect.

Error-seeded ("nocebo")

Realistic data issues planted on purpose and listed in the answer key. Use it to measure what your checks catch and what they miss.

  • sample size
  • allocation ratio
  • true effect size
  • dropout rate
  • AE rate
  • random seed

Who it's for

Statistical programmers

Test SDTM, ADaM and TLF programs against data where the correct output is already known.

Tool and AI builders

Benchmark generators and validators on studies with a scored answer key instead of real patient data.

Trainers and students

Teach the whole pipeline on realistic studies that can be shared freely.

Roadmap

  1. Prototype

    Three therapeutic areas, clean and error-seeded modes, zip export.

  2. More designs

    Crossover, withdrawal, group sequential, non-inferiority and pooled analyses.

  3. More therapeutic areas

    One end-to-end reference study per area.

  4. Open-source release

    Code under Apache-2.0, generated data under CC-BY-4.0.