Case study: synthetic tabular data

Generation is not imputation

Imputation fills missing fields for subjects who already belong to a cohort. Synthetic generation creates new rows whose fields are all generated under a model. The two tasks can support different estimands and fail in different ways. An imputer can be judged by reconstruction of a known masked cell; a generator must be judged by what it preserves, what it distorts, what a downstream model does with the rows, and whether it reproduces information that should not have been exposed.

In an augmentation study, a synthetic row may include a generated outcome so that a downstream model can be trained on complete rows. That is a project design choice, not evidence that a generated outcome is a clinical measurement. The generated table inherits the semantics and distortions of the table used to train it. “Synthetic” describes how a row was produced; it is not a guarantee of realism, usefulness, privacy, or clinical validity.

Two dated model families

The M1-08 case used CTGAN and TabDDPM as dated illustrative implementations for a mixed-type clinical table. CTGAN frames generation as an adversarial problem: a generator proposes rows and a discriminator distinguishes generated from real training rows. Its conditional mechanism can request a category such as a subgroup during generation. The algorithm’s paper describes design choices for continuous and categorical columns; the case’s configuration remains a project-specific use, not a recommendation to use CTGAN for every table.

In that bounded configuration, CTGAN was trained on the complete-case training reference and was asked to generate a specified female deficit. The conditioning was part of the generation mechanism, so the requested subgroup was built into the route rather than selected after generation. The exact count and training composition belong to the dated project record.

Diffusion is a different generation mechanism

TabDDPM represents tabular generation as a diffusion process. A forward process adds noise to numerical and categorical representations; a learned reverse process denoises from noise into a new row. It is not the same adversarial game as CTGAN, and a favourable result for one model family does not establish a general ranking between them.

The M1-08 route used standard unconditional TabDDPM. Its subgroup construction therefore differed from the CTGAN route: rows were generated without a subgroup condition and then inspected in a frozen order until rows with the requested category were found or the declared sampling cap was reached. This is unconditional generation plus selection, not conditional TabDDPM. The distinction matters because selection changes yield and can distort the composition of the retained rows.

The two model names and upstream implementations are dated details. The stable lesson is to describe the generation mechanism, conditioning, selection, and quality-filtering rules precisely enough that a reader can tell what was actually evaluated.

Conditional generation is not generate-then-filter

There are at least two ways to request more rows from a subgroup. A truly conditional generator receives the subgroup condition as part of generation. An unconditional generator can instead produce a stream of rows and keep the first rows meeting the category criterion. The latter may be practical, but it has a measurable yield, a stopping rule, and a selection-induced distribution.

Quality-filtering the generated rows after looking for the most plausible ones would create another selection step. If the goal is to evaluate a generator, cherry-picking rows for fidelity makes the result describe the filter rather than the generator. A protocol should declare whether filtering is permitted, how it is applied, and whether the same rule is applied to every method.

Utility has more than one layer

Utility should be decomposed rather than collapsed into one score. For this chapter, one synthesis keeps four appraisal questions separate rather than presenting a universal taxonomy: marginal fidelity compares distributions of individual continuous and categorical fields; dependency preservation compares relationships among fields, because matching every histogram can coexist with broken associations; downstream utility asks whether a model trained with the synthetic rows performs better, or at least comparably, on untouched real test subjects than a model trained on real data alone; and disclosure risk asks whether the generation process reproduces information that should not be exposed. Kaabachi et al.’s scoping review maps the need for multidimensional utility and privacy evaluation; the four-part arrangement here is a chapter synthesis, not a universal taxonomy.

The last comparison is especially important. Evaluating a downstream model on synthetic rows can reward the generator for reproducing its own assumptions. An untouched real test cohort asks whether the generated data help with the real task. The comparison must also give the real-only and augmented models the same split, fitted-state rules, tuning opportunity, and reporting depth.

The dated M1-08 case kept these families separate. Its aggregate provenance is recorded in the public project record at the pinned commit; the link is provenance for this bounded case, not a general privacy, augmentation, or clinical-validity source. Its downstream survival comparison used one frozen real test cohort. The reference clinical-predictors-only model had C-index 0.789; the TabDDPM- augmented runs ranged from 0.743 to 0.817 across the recorded seeds, with overlapping intervals and no demonstrated improvement. These are bounded project measurements, not a claim about synthetic augmentation in general.

The same case retained negative findings: no generator met the project’s KS-based fidelity criterion, the CTGAN conditional mapping was not demonstrated under that configuration, and memorisation-risk indicators were elevated for TabDDPM on two of three seeds. Reporting all recorded seeds rather than only the most favourable one is part of the evidence boundary.

Risk is not inferred from the word synthetic

A generator trained on a small cohort can reproduce or nearly reproduce training rows. Risk evaluation therefore asks different questions from fidelity and utility. Mixed-type nearest-neighbour distances can compare synthetic-to- training proximity with held-out-real-to-training proximity. Exact matches and near-duplicates can be counted under a declared field representation. Impossible combinations, subgroup distortion, and schema violations can expose structural problems before a row is used downstream.

These are risk indicators, not guarantees of privacy or resistance to a formal disclosure attack. A privacy claim therefore needs a declared threat model and a suitable privacy analysis; proximity indicators alone are not guarantees. A row can be realistic-looking without being useful, private, or clinically valid. Conversely, a generator that avoids close matches can still produce a table whose dependencies are wrong and whose downstream utility is poor. Fidelity, utility, and risk must remain separate claims with separate analyses.

The measurement layer still governs the table

Synthetic generation does not remove the need to define the variables being generated. A clinical field is a measurement with a time, source, coding rule, and missingness process. A derived endpoint may be useful for a bounded project while not being a native semantic of the source dataset. The generator learns whatever relationships the table encodes, including shortcuts, coding artifacts, and errors in those definitions.

This is the tabular version of the measurement-layer argument from Chapter 5. A synthetic row must not be joined to a real patient as if its generated outcome were an observed fact. It is an additional model output whose use must be specified by the study design.

What the dated case can and cannot show

The M1-08 case can show how two declared generation routes behaved under one schema, one split, one augmentation target, one downstream test cohort, and a finite set of seeds. It can show why conditional generation differs from generate-then-filter, why marginal similarity is insufficient, and why risk indicators should be reported alongside utility.

It cannot show that synthetic data are anonymous by construction, that a generator will transport to another cohort, or that a negative augmentation result applies to every model and table. It also cannot replace a clinical validation study. The measured C-index range and risk indicators remain properties of the dated comparison, not universal thresholds or guarantees.

How to read a synthetic-data claim

A self-contained appraisal can proceed without patient rows or private project files. Identify the training reference and held-out real cohort; establish whether generation was conditional, unconditional, filtered, or quality-selected; separate marginal, dependency, downstream, and risk analyses; and check that the downstream test stayed real and untouched. Then ask which seeds, failures, stopping rules, and negative findings were reported.

That reading method carries the chapter’s central boundary into a fast-changing tool area. Algorithm names will age, but the estimand, fitted-state boundary, real-data test, and claim scope remain inspectable. The next chapter begins where this one ends: a validated statistical procedure still depends on images as acquired, located, modality-specific measurements. The medical-imaging AI chapter takes up that measurement-aware model input.

Sources

• Xu L, Skoularidou M, Cuesta-Infante A, Veeramachaneni K. Modeling Tabular Data using Conditional GAN. NeurIPS 2019. arXiv:1907.00503.
• Kotelnikov A, Baranchuk D, Rubachev I, Babenko A. TabDDPM: Modelling Tabular Data with Diffusion Models. ICML 2023, PMLR 202:17564-17579. arXiv:2209.15421 / PMLR.
• Kaabachi B, Despraz J, Meurers T, et al. A scoping review of privacy and utility metrics in medical synthetic data. NPJ Digital Medicine. 2025;8:60. doi:10.1038/s41746-024-01359-3, PMCID PMC11772694.