Functional Decomposition and Shapley Interactions for Interpreting Survival Models

A research team led by Sophie Hanna Langbein introduces Survival Functional Decomposition (SurvFD) and SurvSHAP-IQ. These tools provide a principled framework for interpreting machine learning models that predict time-to-event outcomes, like patient survival. By decomposing higher-order, time-dependent feature interactions, they address the inherent non-additivity of hazard functions. This allows data scientists and medical researchers to understand precisely when and why standard explanation methods fail for critical predictive tasks.