Calibration-Aware Conditional Synthetic Augmentation under Explicit Feasibility Constraints: Repeated Validation on Mobile and Sensor Miscarriage-Risk Records

Synthetic tabular data are increasingly used for augmentation and privacyoriented data access, but their value for probabilistic risk prediction cannot be inferred from distributional resemblance or discrimination alone. We evaluated a calibration-aware conditional synthetic augmentation framework for miscarriage-risk prediction from repeatedly collected mobile and healthcaresensor records. A conditional Wasserstein generative adversarial network with gradient penalty was conditioned on the binary outcome and prespecified age and body-mass-index strata. Its objective combined adversarial learning with moment matching, correlation preservation, raw-domain feasibility, and label-consistency penalties. Generative and predictive hyperparameters were selected on a dedicated tuning partition using constrained Bayesian optimization and then locked for 20 repeated held-out evaluations. In each repetition, elastic-net logistic models were fitted to real records alone or to real plus 4,000 synthetic records. Recalibration was estimated only from a separate real-data calibration partition, and all metrics were evaluated on held-out real records. Synthetic augmentation did not change discrimination: mean AUC was 0.641 for both training strategies, with a paired difference of 0.000 (0.000, 0.000). It produced small improvements in Brier score (-0.001), log loss (-0.002), absolute calibration-slope error (-0.248), and 10-bin equal-width ECE (-0.071). Tie-preserving 10-bin ACE decreased by -0.007, but its resampling interval included zero (-0.016 to 0.002; adjusted Wilcoxon p = 0.751). Differences largely disappeared after Platt or temperature calibration. No Bayesian-optimization candidate satisfied all feasibility constraints; mean classifier two-sample-test AUC was 0.612, mean pre-projection domain-violation rate was 87.5%, and the nearest-neighbour memorization proxy was 0.022%. Conditional augmentation therefore acted as a modest regularizer for the uncalibrated predictor, but it neither improved ranking nor yielded a generator that naturally respected the prespecified feature domain. Because the data comprise repeated records and were split at record level, the findings are methodological and do not establish clinical deployment readiness.