JMIR Publications

ABSTRACT

Background:

Sharing patient-level data for downstream classifier training is constrained by privacy risks and by group-level disparities that synthetic data can either inherit or amplify. Fidelity-oriented tabular generators optimize distributional match and downstream prediction accuracy without explicit fairness consideration, whereas fairness-aware generators add a parity penalty during training that competes for the same parameters as the reconstruction loss. The two objectives are therefore not separated, and mechanisms that apply fairness at sampling time rather than inside the training loss have not been systematically evaluated on clinical cohorts.

Objective:

To develop a two-stage synthetic tabular data generator, FairVQD, that decouples reconstruction and fairness by applying the fairness intervention at sampling time rather than during training, and to evaluate its utility–fidelity–fairness trade-off on three patient cohorts spanning small-imbalanced, large-balanced, and large-high-cardinality regimes.

Methods:

FairVQD combines a vector-quantized autoencoder trained for reconstruction only (Stage 1) with an input-space denoising diffusion refiner trained under the elucidating diffusion models (EDM) formulation with classifier-free guidance (Stage 2). At sampling time, the binary protected attribute is drawn from a uniform Bernoulli distribution rather than from the biased data marginal, so the synthetic marginal is exactly balanced. We evaluated FairVQD on DiaBD, the Cardiovascular Disease, and Diabetes-130 datasets. We compared against tabular generators: Gaussian Copula, Conditional Tabular Generative Adversarial Network (CTGAN), and Tabular Variational Autoencoder (TVAE). We measured fidelity using column-shape and pair-correlation errors; downstream utility using the train-on-synthetic, test-on-real area under the receiver operating characteristic curve (TSTR-AUC); and fairness using the demographic parity ratio (DPR) and equalized odds ratio (EOR).

Results:

Across the three cohorts, FairVQD achieved the highest mean EOR (0.90, averaged over Cardiovascular and Diabetes-130; the DiaBD test set contained only 34 positives, which collapsed the EOR ratio) and the highest mean DPR (0.81) among all evaluated methods. Against the strongest non-fair baseline, CTGAN, mean EOR was 38% higher (0.90 vs 0.65) and mean DPR was 5% higher (0.81 vs 0.77). The mean downstream TSTR-AUC was 0.65, which was 4 percentage points below CTGAN (0.69) and 9 points below TVAE (0.74). Mean column-shape error was 0.30 for FairVQD compared with 0.08 for both CTGAN and TVAE, and mean pair-correlation error was 0.46 compared with 0.28–0.29. The downstream-AUC cost was concentrated on the Cardiovascular cohort (12 points against both CTGAN and TVAE); on DiaBD and Diabetes-130, FairVQD matched or near-matched CTGAN. An architecture ablation removing Stage 2 collapsed both fairness ratios to near zero (mean DPR 0.06, mean EOR 0.05) and degraded fidelity.

Conclusions:

Applying fairness at sampling time via classifier-free guidance over a uniformly drawn protected attribute decouples fidelity from fairness during training and yields higher parity than fidelity-only baselines across three clinical regimes, at an accounted-for cost in downstream accuracy..