GRAFT: Decoupling Ranking and Calibration for Survival Analysis

In three linesGRAFT is a hybrid AFT model for survival analysis that decouples prognostic ranking from calibration of survival estimates. It combines a linear AFT model with a non-linear residual neural network and stochastic gates for feature selection. Trained on C-index-aligned ranking loss with conditional imputation, it outperforms baselines in discrimination and calibration.

Summary generated by Claude — human-verified