source: arxiv machine learning: a filtered mixture-of-generators for fully synthetic survival training
level: research
survival analysis models time-to-event data but clinical training data is scarce. events take years to accumulate, cohorts are small, and privacy rules block data sharing. tabular generative models can create synthetic data for augmentation or privacy-safe sharing, but they need large datasets to work well. on small survival cohorts, a single generator often fails to capture the population, so models trained on its output do not match real-data performance.
fogs (filtered mixture-of-generators for survival analysis) treats synthetic data creation as sample selection instead of generation. it draws a candidate pool from four different tabular generators. each sample is scored by an ensemble of seven survival models trained on real data. the scoring uses proper scoring rules as a per-sample plausibility measure. a two-level pipeline then optimizes the selection to build a high-quality synthetic training set.
the approach avoids relying on one generator and instead picks the most realistic samples from a diverse pool. this helps downstream survival models trained on the synthetic data perform closer to those trained on real data. the method is designed for small clinical cohorts where standard generative models struggle. it offers a practical way to augment limited data or share synthetic cohorts without compromising privacy.
why it matters: it enables better survival model training on small, privacy-sensitive clinical datasets by selecting high-quality synthetic samples from multiple generators.
source: arxiv machine learning: a filtered mixture-of-generators for fully synthetic survival training