Support Vector Machine for Dynamic Survival Prediction with Time-Dependent Covariates

• Activity Number: 82 - New Statistical Methods for Survival Analysis in Complex Biomedical Studies
• Type: Invited
• Date/Time: Monday, August 8, 2022 : 8:30 AM to 10:20 AM
• Sponsor: Lifetime Data Science Section
• Author(s): Wenyi Xie and Donglin Zeng* and Yuanjia Wang
• Companies: University of North Carolina and University of North Carolina and Columbia University
• Keywords: Counting process; Survival prediction; Dynamic prediction; Time-dependent covariates; Machine learning

Abstract

Many machine learning approaches such as random forest and support vector regression have been proposed for survival prediction but most of them are restricted to utilizing only baseline covariates. On the other hand, time-evolving biomarkers can provide more timely and accurate information for future prediction. To incorporate these biomarkers into the prediction, we propose a new framework based on support vector machines for the counting process for the survival time. We show theoretically that our method is equivalent to comparing the time-sensitive hazards rates among at-risk subjects and we further obtain the convergence rate of the resulting prediction function. Simulation studies and application to a study on Huntington's Disease are provided to compare the numerical performance between the proposed method and existing machine learning approaches.