Treatment Effect Estimation and Therapeutic Optimization Using Observational Data

Abstract

In this dissertation, I address two essential questions of modern therapeutics: (1) to quantify the e ects of pharmacological agents as functions of patient's clinical characteristics; (2) to optimize individual treatment regimen in the presence of multiple treatment options. To address the rst question, I proposed a uni ed framework for the estimation of heterogeneous treatment e ect (x), which is expressed as a function of the patient characteristics x. The proposed framework not only covers most of the existing advantage-learning methods in the literature, but also enhances the robustness of di erent learning methods against outliers by allowing the selection of appropriate loss functions. To cope with high-dimensionality in x, I incorporated into the method modern machine learning algorithms including random forests, gradient boosting machines, and neural networks, for a more scalable implementation. To facilitate the wider use of the developed methods, I developed an R package RCATE, which is now posted on Github for public access. For therapeutic optimization, I developed a treatment recommendation system using o ine reinforcement learning. O ine reinforcement learning is a type of machine learning method that enables an agent to learn an optimal policy in the absence of an interactive environment, such as those encountered in the analysis of therapeutics data. The recommendation system optimizes long-term reward, while accounting for the safety of treatment regimens. I tested the method using data from the Systolic Blood Pressure Trial (SPRINT), which included multiple years of follow-up data from thousands of patients on many di erent antihypertensive drugs. Using the SPRINT data, I developed a treatment recommendation system for antihypertensive therapies.