Synthetic data generation from population-based breast cancer registries: opportunities and limitations for survival...

Synthetic data generation via Generative Adversarial Networks (GANs) has emerged as a promising strategy for privacy-preserving data sharing, cohort augmentation, and synthetic control arm construction in oncology. However, the extent to which GAN-derived cohorts preserve survival dynamics alongside covariate structure remains poorly characterised. This study evaluated structural fidelity, survival concordance, and prognostic preservation in a large synthetic breast cancer cohort derived from a population-based registry.

A Conditional Tabular Wasserstein GAN (CT-WGAN) was trained on 325,896 women with invasive breast cancer from the SEER-17 database (2004-2012). Structural fidelity was quantified using the SAFE framework (Distance Score, Correlation Score, PCA Score; composite Clinical Synthetic Fidelity [CSF] index). Survival concordance was assessed through Kaplan-Meier analysis, log-rank testing, and hazard ratio comparison for overall survival (OS) and breast cancer-specific survival (BCSS). Prognostic concordance was evaluated by independent multivariable Cox models in real and synthetic cohorts.

The synthetic cohort achieved high structural fidelity (CSF = 0.916; Distance Score 0.932, Correlation Score 0.949, PCA Score 0.868). Despite this, survival events were substantially underrepresented (OS events 13.8% vs 36.8%; BCSS mortality 4.6% vs 11.5%), with markedly superior Kaplan-Meier estimates in the synthetic population (OS HR synthetic vs real 0.369, p < 0.0001). Directional prognostic concordance was preserved for four of seven variables; ER status lost its protective effect in the synthetic cohort.

Structural fidelity and survival concordance are distinct validation dimensions. GAN-derived cohorts are suitable for covariate-level applications but require dedicated outcome recalibration before use in survival analyses or synthetic control arm construction.