A0759
Title: Optimal dynamic treatment regimes via smooth surrogate losses
Authors: Aaron Sonabend - Google Research (United States)
Nilanjana Laha - Texas A&M (United States) [presenting]
Rajarshi Mukherjee - Harvard T.H. Chan School of Public Health (United States)
Tianxi Cai - Harvard School of Public Health (United States)
Abstract: Large health care data repositories such as electronic health records (EHR) open new opportunities to derive individualized treatment strategies for complicated diseases. We will discuss the problem of estimating sequential treatment rules tailored to a patient's individual characteristics, often referred to as dynamic treatment regimes (DTRs). Our main objective is to find the optimal DTR that maximizes a discontinuous value function through direct maximization of Fisher's consistent surrogate losses. We demonstrate that a large class of concave surrogates fails to be Fisher consistent -- a behavior that differs from traditional binary classification problems. We further characterize a non-concave family of Fisher consistent smooth surrogates that can be optimized via gradient descent using off-the-shelf machine learning algorithms. Compared to the existing direct search approaches under the support vector machine framework, our proposed method is more computationally scalable to large sample sizes and allows for broader functional classes for treatment policies. We establish theoretical properties for our proposed DTR estimator and obtain a sharp upper bound on the regret. The finite sample performance of our proposed estimator is evaluated through extensive simulations. Finally, we illustrate the working principles and benefits of our method using EHR data from sepsis patients admitted to intensive care units.
