B1356
Title: Dynamic risk prediction with rank-based survival trees
Authors: Yifei Sun - Columbia University (United States) [presenting]
Mei-Cheng Wang - Johns Hopkins University (United States)
Abstract: Tree-based methods are popular statistical tools for creating simple and interpretable prediction rules with mild assumptions. We introduce a unified framework for tree-structured analysis using binary response or survival data. With arguments from the Neyman-Pearson Lemma, we propose rank-based methods for growing and pruning trees guided by a concordance index. We define the concordance index as a map from an arbitrary function of the covariates to a real number, where the target function maximizes the concordance. In contrast with the existing methods where each split maximizes between node heterogeneity or within node homogeneity, our approach aims to maximize the concordance index of the whole tree. For right-censored survival data, our framework has the flexibility to incorporate time-dependent covariates, resulting in more accurate prognostic models than only considering baseline covariates.
