New Statistical Insights to Precision Medicine, from Targeted Treatment Development to Individualized Tailoring Recommendation

There has been increasing interest in discovering precision medicine in current drug development. One aspect of precision medicine is to develop new therapies that target a subgroup of patients with enhanced treatment efficacy through clinical trials. Another aspect is to tailor existing therapies to each patient so that everyone can get the most “suitable” treatment. Motivated by analyzing the Age-Related Eye Disease Study (AREDS) data, this dissertation proposes new statistical methods to address issues in both aspects. In the first part, I propose a novel multiple-testing-based approach to simultaneously identify and infer subgroups with enhanced treatment efficacy. Specifically, I formulate the null hypotheses through contrasts and construct their simultaneous confidence intervals, which control both within- and across-marker multiplicity. Two types of outcomes are considered: survival and binary endpoints. Extensive simulations are conducted to evaluate the method performance and provide practical guidance. The method is then applied to AREDS data to assess the efficacy of antioxidants and zinc in delaying AMD progression. I further validate the findings in AREDS2, by discovering consistent differential treatment responses in subgroups identified from AREDS. In the second part, I develop machine-learning-based approaches to estimate individual treatment effects (ITE) so that individualized tailoring recommendation can be provided. Specifically, I implement random survival forest, Bayesian accelerated failure time model, and Cox-based deep neural network survival model under the framework of meta-algorithms: T-learner and X-learner, to accurately estimate ITEs with survival outcomes. Treatment recommendation rule is provided based on patient’s ITE estimate and then evaluated by various performance metrics. I investigate the merits of the proposed methods with comprehensive simulation studies and apply them on AREDS data. Finally, the Boruta algorithm is applied to identify top variables that contribute to the treatment recommendation rule. Public health significance: This dissertation addresses two precision medicine research questions: (1) targeted treatment development, i.e., whether there exists subgroup of patients with beneficial treatment efficacy; (2) tailoring existing therapies through ITE estimation. It has the potential to significantly improve the current practice in analyzing treatment effects, and thus to increase the success of modern drug development and precision medicine research