부산대학교 도서관

Summary: Optimal dynamic treatment regimen (DTR) is one of the most important strategies in precision medicine, which sequentially assigns the best treatment to patients based on their evolving health status to maximize the cumulative outcome. For many chronic diseases, treatments are often multifaceted where aggressive treatments with a higher beneficial reward are usually accompanied by an elevated risk of adverse outcomes, and ideal DTRs should both yield a higher beneficial gain while avoiding unnecessary risk. In addition, it is often that among many possible tailoring variables, only a small subset is essential for treatment, and identifying these variables is particularly important for developing sparse DTRs, which are useful in practice. To address these challenges, in the first project we propose a new machine learning-based method to learn the optimal DTRs that maximize patients' cumulative reward but at each stage, the acute short-term risk induced by the treatments is controlled lower than a pre-specified threshold. We show that this multistage-constrained problem can be decomposed into a series of single-stage single-constrained problems, which can be efficiently solved using a backward algorithm. We provide theoretical guarantees for the method and demonstrate the performance via simulation studies and an application to a clinical trial for T2D patients (DURABLE study). In the second project, we develop a general approach to estimate the optimal DTRs that maximize patients' cumulative reward but lead to a cumulative risk no higher than a pre-specified threshold. This procedure converts the problem into solving unconstrained DTRs problems, which can be accommodated to existing DTRs methods. Furthermore, we propose an estimation procedure (MRL) to solve the decision rules across all stages simultaneously. The method is justified via theoretical guarantees, simulation studies, and an application to the DURABLE study. In the third project, we develop a new machine learning-based method by extending and adding an L1-penalty to the MRL framework to implement variable selection while learning optimal DTRs across all stages contingently. A DC algorithm is developed to solve the L1-MRL problem efficiently and the performance is demonstrated via simulation studies and application to an observational electronic health record (EHR) data of T2D patients.