Synthetic Data Filtering Boosts Survival Model Training

Survival analysis, which models time-to-event data, faces significant challenges in clinical settings due to the scarcity and high cost of training data. Events accrue over long periods, cohorts are often small, and strict privacy regulations limit data sharing across institutions. While tabular generative models offer potential solutions for data augmentation and privacy-preserving sharing, a single generator typically struggles to adequately characterize small cohorts, leading to suboptimal performance when downstream models are trained on its synthetic output. FoGS (Filtered Mixture-of-Generators for Survival analysis) addresses this by re-conceptualizing synthetic data construction as a sample selection problem. Instead of relying on a single generator, FoGS creates a candidate pool from four architecturally distinct tabular generators. Each synthetic sample is then scored for plausibility by an ensemble of seven survival models, which were initially trained on real data, using proper scoring rules. The method employs a two-level optimization pipeline. The outer loop optimizes a selection policy, including generator quotas, scorer weights, a random complement, and stratified balancing on event time and censoring, against the performance of a held-out downstream model (XGBoost-Cox). The inner loop tunes this downstream model. Experiments on 16 public datasets show that FoGS significantly improves C-index and IBS metrics when training on synthetic data and testing on real data, often matching or exceeding real-data training performance without compromising privacy margins.