Tree-Derived Survival Risk Groups in Differentiating Care for Glioma Patients

• Activity Number: 6
• Type: Invited
• Date/Time: Sunday, August 3, 2014 : 2:00 PM to 3:50 PM
• Sponsor: WNAR
• Author(s): Annette M. Molinaro*+ and Robert Strawderman and Adam Olshen
• Companies: University of California, San Francisco and University of Rochester and University of California, San Francisco
• Keywords: Loss function ; Tree ; Risk prediction ; partDSA ; Genomics ; Glioma

Abstract

We recently developed partDSA, a multivariate method that, similarly to CART, utilizes loss functions to select and partition predictor variables to build a tree-like regression model for a given outcome. However, unlike CART, partDSA permits both 'and' and 'or' conjunctions of predictors, elucidating interactions between variables as well as their independent contributions. partDSA thus permits tremendous flexibility in the construction of predictive models and has been shown to supersede CART in both prediction accuracy and stability. As the resulting models continue to take the form of a decision tree, partDSA also provides an ideal foundation for developing a clinician-friendly tool for accurate risk prediction and stratification.

With right-censored outcomes, partDSA builds estimators via either the Inverse Probability Censoring Weighted (IPCW) or Brier Score weighting schemes; see Lostritto et al. (2012), where it is shown that partDSA performs as well or better than two competing tree-based methods. In this talk, useful extensions of partDSA are described and we show the power of partDSA in deriving survival risk groups for glioma patients based on genomic markers.