FoGS: New AI generates synthetic survival data beating real training

FoGS (Filtered Mixture-of-Generators for Survival analysis) reframes synthetic-data construction as sample selection rather than generation, addressing a critical bottleneck in clinical machine learning: survival analysis datasets are often small, expensive to collect, and privacy-restricted. The method first creates a candidate pool using four architecturally distinct tabular generators (e.g., CTGAN, TVAE, etc.). Each generated sample is scored by an ensemble of seven survival models trained on real data, using proper scoring rules as a per-sample plausibility proxy. A two-level optimization pipeline then selects the best combination of generator quotas, scorer weights, a random complement, and stratified balancing on event time and censoring, with an inner loop tuning the downstream XGBoost-Cox model.

Evaluated on 16 public datasets under a train-on-synthetic, test-on-real paradigm (using C-index and Integrated Brier Score on a 0-100 scale), FoGS achieved mean improvements of +2.17 in C-index and +0.67 in IBS. It improved both metrics on 9 of 16 datasets and at least one on 13, with statistical significance (Wilcoxon p=0.039 and p=0.035). In most cohorts, synthetic training matched or exceeded real-data performance, and privacy analysis showed no significant increase in nearest-neighbour disclosure risk compared to unfiltered sampling. This makes FoGS a practical tool for privacy-preserving survival analysis without sacrificing predictive power.