Methods for generating synthetic data

Family 01 · Rules and simulation

Knowledge-driven simulation

You write down what you believe about the population and the trial, then let the computer generate patients. Statisticians have done this for decades to check sample sizes and trial designs. Examples include a two-arm trial with assumed event rates and dropout, PK/PD models linking dose to concentration and response, and Synthea, which simulates patient records along care pathways.

A two-arm survival trial in about ten lines of Python. Every number is an assumption you can change. Try it interactively in the lab →

import numpy as np, pandas as pd
rng = np.random.default_rng(2026)

n = 300                              # patients per arm
median_ctrl, hr = 6.0, 0.75          # assumptions (months, hazard ratio)
lam_c = np.log(2) / median_ctrl      # exponential hazard, control
arm   = np.repeat(["Placebo", "Active"], n)
lam   = np.where(arm == "Active", lam_c * hr, lam_c)
t_evt = rng.exponential(1 / lam)     # time to event
t_cen = rng.uniform(6, 24, 2 * n)    # administrative censoring
df = pd.DataFrame({"arm": arm,
     "time":  np.minimum(t_evt, t_cen),
     "event": (t_evt <= t_cen).astype(int)})

Family 02 · Resampling

Resampling and perturbation

Bootstrap draws real records with replacement. Noise addition makes small random changes so no record is an exact copy. SMOTE-type methods create new records between existing ones, often to boost rare classes.

Family 03 · Statistical models

The Gaussian copula in three steps

1. Learn each column. What does age look like on its own? Tumour size? Learn the shape of every variable.
2. Learn how they move together. Older patients have more comorbidities. Learn these correlations separately.
3. Generate new rows. Draw correlated random numbers and map them back onto each column's shape.

Family 03 · Statistical models

Build a patient one variable at a time

Bayesian networks and sequential CART (for example the R package synthpop) generate each variable from the ones already generated, much like following a clinical story.

It handles mixed data types and nonlinear relationships well, lets you build clinical logic into the variable order, and works with a few hundred to a few thousand records. The order matters, and very many variables make it slow.

library(synthpop)
syn_obj <- syn(real_df, method = "cart", seed = 2026)
synthetic_df <- syn_obj$syn
compare(syn_obj, real_df)   # side-by-side distributions

Family 04 · Deep learning

Why deep learning?

Deep generators help when there are complex, interacting relationships, hundreds of variables, sequences over time such as visits and treatment histories, or unstructured text and images.

Family 04 · Deep learning

GANs: a forger and a detective

A generator (the forger) turns random noise into fake patients. A discriminator (the detective) tries to tell fake from real. Each round of feedback makes both better, and training stops when the detective can no longer tell the difference.

Tabular versions for mixed clinical data include CTGAN, CopulaGAN and Wasserstein GANs such as WGAN-GP and conditional tabular WGANs.

Family 04 · Deep learning

Three more families worth knowing

Clinical reality

What makes clinical data hard to synthesise

Case study

Synthetic data from a lung cancer trial

Using a randomised lung cancer trial (Tarceva versus placebo) with overall survival and progression-free survival endpoints, provided by RSS under a data-sharing agreement, we compared a Gaussian Copula Synthesizer with a conditional tabular Wasserstein GAN.

Both were judged with the same checks: Kaplan–Meier curves, hazard ratios, distributions and closeness to real records.

Software

You do not need to code these from scratch

From real table to synthetic table in a few lines with SDV. Running the generator is the easy part; metadata, rules and evaluation are where the effort goes.

from sdv.metadata import Metadata
from sdv.single_table import GaussianCopulaSynthesizer

metadata = Metadata.detect_from_dataframe(data=real_df)
synth = GaussianCopulaSynthesizer(metadata)
synth.fit(real_df)
synthetic_df = synth.sample(num_rows=500)

# Try a GAN by changing one line:
# from sdv.single_table import CTGANSynthesizer
# synth = CTGANSynthesizer(metadata, epochs=300)

Evaluation

Three questions for any synthetic dataset

These pull against each other: the closer the data are to the real data, the higher the privacy risk. The purpose decides the balance. See the trade-off in the lab →

Evaluation

Practical checks

Privacy

Synthetic does not automatically mean anonymous

Measure exact and near copies (distance to the closest real record), test whether an attacker could tell if a person was in the training data (membership inference), and look for rare combinations that point to one person. Differential privacy adds mathematical guarantees, usually at some cost to accuracy.

Pitfalls

Five common mistakes