Xiaogang Su ColloquiumThursday, October 20, 2016 - 2:45pm
Statistics Department Colloquium
Where: LeConte College, Room 210
Speaker: Xiaogang Su
Title: A Look into Personalized Medicine via Interaction Trees
There has been a growing interest in personalized medicine which essentially involves assessing heterogeneous treatment effects. Concerning experimental data collected from randomized trials, we explore stratified and individualized treatment effects with a machine learning approach -- Interaction Trees (IT; Su et al., 2009). We first propose a smooth sigmoid surrogate (SSS) splitting method, as an alternative to greedy search (GS), to speed up GS and amend its deficiencies. On the basis of modified IT, causal inference at different levels can be made. More specifically, an aggregated grouping procedure stratifies data into refined subgroups where the treatment effect remains homogeneous in each. Ensembles of IT models can provide prediction for individualized treatment effects (ITE), which compares favorably to the traditional ‘separate regression’ methods. In order to extract meaningful interpretations, we have also made available several other features such as variable importance ranking, partial dependence plot to help identify important effect moderators for the treatment among high-dimensional covariates, and ensemble majority voting for determining the optimal treatment regime. An empirical illustration of the proposed techniques is made via an analysis of quality of life (QoL) data from breast cancer survivors.
Xiaogang Su Colloquium Flyer