Post date: Apr 12, 2018 4:47:58 PM
Short Course
Title: `Data-driven methods for subgroup identification and personalized medicine'
Lecturer: Dr. Jared Foster, National Cancer Institute
Abstract: In this talk, I will present an introduction to some topics within personalized medicine, and will discuss a variety of methods for subgroup identification. I will primarily focus on settings in which the ultimate goal is to identify subgroups, across which the expected treatment effect differs (should they exist). I will present a general introduction to a few different frameworks within which this problem can be approached, and for each, will briefly discuss a few specific examples of methods, some of which I will illustrate using real data. In the implementation of these methods, a variety of machine learning techniques will be considered, including CART, random forests, and (LASSO) penalized single-index models. I will begin by discussing tree-based methods, which work by recursively partitioning the data into subgroups, and which can be used to attack a variety of subgroup identification problems, depending on how specifically the data are partitioned. Another, more general way to approach this problem is to model the conditional expected treatment effect as a function of the covariates. This can be a very useful way to identify subgroups, and more generally, can help one to better understand the extent to which the expected treatment effect varies with the patient characteristics. Finally, I will briefly introduce the framework of optimal treatment regimes (OTR), and will present a few specific ways to approach subgroup identification within this framework. OTR essentially turns the subgroup identification problem into an optimization problem, where the goal is now to identify the “best” subgroup among a class of possible subgroups. The idea here is that subgroup membership will be used to assign treatment for future patients, and thus we can look at these subgroups as defining treatment regimes.