Survival Trees and Forest for Thyroid Cancer Prognostication

• Activity Number: 409
• Type: Topic Contributed
• Date/Time: Tuesday, August 6, 2013 : 2:00 PM to 3:50 PM
• Sponsor: International Indian Statistical Association
• Author(s): Mousumi Banerjee*+ and Daniel Muenz and Megan Haymart
• Companies: University of Michigan and University of Michigan and University of Michigan
• Keywords: censored data ; tree-based methods ; splitting rules ; forest ; thyroid cancer

Abstract

Tree-based methods partition the covariate space into a set of rectangles, leading to a fitted model that is piecewise constant over regions of the covariate space. Trees are conceptually simple yet powerful, and are being increasingly used in biomedical studies for analyzing censored survival data where the primary goal is prognostication of patients. This talk gives an overview of the methodological aspects of tree-based modeling, with a focus on comparing and contrasting different splitting criteria in the survival data setting. Splitting rules that use within-node error versus between-node separation measures are compared using simulated and real data. To gain accuracy in prediction and address instability in a single tree, we present an extension of survival trees to an ensemble of trees. The methods are illustrated using data from the US National Cancer Database to model survival of patients with thyroid cancer. Although thyroid cancer is a disease that has seemingly good prognosis, using tree-based methods we are able to identify subtypes that have poor prognosis and may benefit from earlier therapeutic intervention.