Methods

TSTR (train on synthetic, test on real)

Also called: TSTR, TRTR

TSTR measures the usefulness of synthetic data: train a model on synthetic records, test it on real ones, and compare with a model trained on real data (TRTR).

Train on synthetic, test on real (TSTR) is a utility test for synthetic data. A predictive model is trained only on the synthetic dataset and evaluated on held-out real data. The same model trained on real data and tested on real data (TRTR) gives the benchmark.

If the TSTR score, often an AUC, is close to TRTR, the synthetic data preserved the relationships needed for that prediction task. A large gap means important structure was lost.

TSTR tests one task at a time, so it is usually reported together with distribution-level fidelity measures and privacy tests.

Tools that use this

EHRSynthUpload a clinical CSV, train a stage-conditional diffusion model and generate privacy-preserving synthetic patients, with fidelity, privacy and TSTR evaluation.

Related terms

Educational reference. Not medical advice and not for clinical decision-making.