A1266
Title: The optimal dynamic treatment regime using smooth surrogate losses
Authors: Nilanjana Laha - Texas A\&M University (United States) [presenting]
Abstract: Large health care data repositories, such as electronic health records (EHR), present new possibilities for deriving personalized treatment strategies for complex diseases like sepsis. This paper focuses on estimating dynamic treatment regimes (DTRs), which are sequential treatment rules tailored to individual patient characteristics. The main objective is to find the optimal DTR by maximizing a discontinuous value function through Fisher-consistent surrogate loss functions. Unlike classical binary classification problems, it is shown that a large class of concave surrogates fails to be Fisher consistent. However, a non-concave family of Fisher consistent smooth surrogate functions is identified, enabling efficient optimization using gradient-descent algorithms. The proposed DTR estimation via surrogate loss optimization (DTRESLO) method outperforms existing direct search approaches in terms of computational scalability for large sample sizes and accommodating broader functional classes for treatment policies. Theoretical properties and a sharp upper bound on the regret for DTRESLO are established. Extensive simulations confirm the finite sample performance of the proposed estimator. The method is illustrated by estimating an optimal DTR for sepsis treatment using EHR data from sepsis patients in intensive care units.