Fully synthetic
Every record is generated. Strongest privacy, but relies most on the quality of the generator.
One-day course for clinical, regulatory and HTA professionals
Data that behaves like patients.
…without being anyone. Synthetic data can speed up trial design, open up access to evidence and protect privacy. This companion site walks through how it is made, how to check it, and how to choose a method you can defend.
Session 2: Methods for generating synthetic data
Dr Saqib Ur Rehman. Three parts of about 25 minutes: the methods toolbox, deep learning generators, and how to judge whether the data are good enough.
Read the session content →
4method families
2live simulations
15practice questions
Try it before the theory
Which points are real?
Purpose→Method→Evaluate→Document
What synthetic data is, and is not
Patterns of real data, records of no one
Artificially generated data that reproduces the patterns and statistical properties of real data, without being a record of any real person.
Partially synthetic
Only the sensitive variables are replaced. The rest of the record stays real.
Simulated
Generated from assumptions or models, such as event rates or PK/PD equations, without learning from patient data.
Three common misconceptions
“It is fake, so it is useless.”
It is useful when it reproduces the patterns a task needs, and that can be tested.
“It is always anonymous.”
Privacy must be measured. Poorly generated data can contain near-copies of real patients.
“It will replace clinical trials.”
It complements trials by supporting design, comparison and extrapolation.
02Purpose first
Start with one question
What must the data do?
The purpose decides how realistic the data must be, which method to use and how you must check it. The bars show how much realism each purpose needs.
Test and train
Build and test code and pipelines without touching patient data.
Realism needed: low
Design and simulate
Explore trial designs, sample sizes and what-if scenarios before enrolment.
Realism needed: medium
Answer questions
Reproduce the real analysis, such as a hazard ratio, so relationships must be preserved.
Realism needed: high
Share safely
Give collaborators data that behaves like the original while protecting individuals.
Realism needed: high + privacy checks
03Experiment room
Interactive simulation lab
Test an idea before trusting a method
Two experiments run in your browser: simulate a whole trial from assumptions, then compare three ways of generating synthetic patients.
Featured experiment
When does realistic become too real?
Switch between independent columns, a copula and a bootstrap, and watch fidelity rise while the share of near-copies of real patients climbs with it.
Run the experiment →Assume → Simulate → Compare → Decide
Change one setting at a time and read what responds.
04Method families
Four families of methods
Start on the left. Move right only when you must.
Move to a more flexible family only when the simpler method is not good enough for your purpose.
FAMILY 01No data needed
Rules and simulation
Build data from assumptions and known models.
Read more →FAMILY 02Quick
Resampling
Reuse and perturb real records: bootstrap, noise, SMOTE.
Read more →FAMILY 03Best baseline
Statistical models
Learn distributions and correlations: copulas, Bayesian networks, CART.
Read more →FAMILY 04Complex data
Deep learning
Learn complex patterns: GANs, VAEs, diffusion, transformers.
Read more →05Course day
The programme
From fundamentals to future directions in one day
Timings are approximate and may shift on the day.
This site covers
Session 2: Methods for generating synthetic data
10:45 to 12:00 with Dr Saqib Ur Rehman. Everything in the session, plus simulations, a method chooser and practice questions, is on this site.
Go to the session content- Registration and networking
- Fundamentals of Synthetic DataDr Ifty Khan
- Break
- Methods for Generating Synthetic DataDr Saqib Ur Rehman. Session content on this site
- Lunch
- Regulatory Perspectives on Synthetic DataMHRA perspective, with a response from RSS
- Applications of Synthetic Data: Case Studies using EclipticaIntroduction to Ecliptica, download and install; using Ecliptica for clinical trial simulation & PK/PD; external control arms
- Break
- Applications of Synthetic DataReal World Data: target trial emulation & digital twins example; Health Technology Assessment for JCA
06Decision corner
Method chooser
Not sure which method fits your project?
Answer four short questions about your purpose and your data. You get a starting method, alternatives, tools and the checks you will need.
A starting methodWhat else to tryThe checks you need
Open the method chooser