Online Program

Random Survival Forests for Dynamic Prediction of a Time-to-event outcome using a Longitudinal Biomarker (309808)

Kristen Campbell, University of Colorado 
Elizabeth Juarez-Colunga, University of Colorado 
Kaci Pickett, University of Colorado 
*Krithika Suresh, University of Colorado 

Keywords: predictive modeling, joint modeling, longitudinal data, machine learning, survival analysis

Risk prediction models for time-to-event outcomes play a vital role in personalized decision-making. Patient biomarkers are often collected at multiple timepoints during a patient’s follow-up, but traditional survival prediction models ignore their longitudinal nature and use only baseline information. Dynamic prediction incorporates longitudinal information to produce updated, patient-specific predictions during follow-up. Existing methods for dynamic prediction include joint modeling, which can suffer from computational complexity and poor performance under misspecification, and landmarking, which is easily implemented in software but traditionally uses a Cox model that relies on the proportional hazards assumption. Random survival forests, a machine learning algorithm for time-to-event outcomes, can capture complex data relationships and has been shown to have superior predictive performance in such situations. We propose an alternative approach for dynamic prediction using random survival forests in a landmarking framework. With a simulation study, we compare the predictive performance of this proposed method to Cox landmarking and joint modeling and find that the performance of RSF landmarking is superior when there exists a complex relationship between the survival and longitudinal marker, and when there are multiple predictors present for which the clinical relevance is unknown. We illustrate the use of the RSF approach in a clinical application to assess the performance and necessity of various RSF model-building decisions, such as imputation and choosing tuning parameters. We conclude that RSF landmarking is a nonparametric, machine learning alternative to current methods for obtaining dynamic predictions when there exist complex or unknown relationships. It requires little prior specification, has comparable predictive performance and computational speed, and can easily accommodate multiple longitudinal biomarkers for predicting survival during follow-up.