Researchers built SurvICL, a machine-learning model that predicts when an event will happen from data where some outcomes are only partly observed. Trained purely on synthetic data, it outperformed the fifty-year-old Cox model and 20 other methods across 63 benchmark datasets, without any tuning.
Survival analysis often works with small patient groups where many outcomes are only partly known. Until now, no newer AI model had outperformed the fifty-year-old Cox method, especially on small datasets. SurvICL changes that with no tuning, which could sharpen risk estimates in exactly those data-scarce clinical settings.
✦ A fifty-year-old standard met its match — a model trained on data no patient ever provided.