New Statistical Learning Methods For Optimizing Personalized Dose

Abstract

As precision medicine and personalized health care continues to revolutionize the field, most existing work still focus on categorical treatment scenarios, instead of continuous treatment options. In addition, the frequently used algorithms such as black-box machine learning methods and tree learning methods are lack of interpretability and global optimality respectively. Many scenarios also need to consider multiple competing healthcare priorities into account. This dissertation in- troduces new optimal dose finding methods to address these challenges, aiming to significantly contribute to the advancement of precision medicine development. In Chapter 2, we propose a non-greedy optimization method for dose search, namely Global Optimal Dosage Tree-based learning method (GoDoTree), which combines a robust estimation of the counterfactual mean outcome with an interpretable and non-greedy decision tree search for estimating the global optimal dynamic dosage decision rules in a multiple treatment stage setting. GoDoTree can estimate how the counterfactual outcome mean depends on a continuous treatment dosage using doubly robust estimators and optimize the decision tree in a non-greedy way. In Chapter 3, we present a novel algorithm for developing individualized treatment regimes (ITRs) that incorporate continuous treatment options and multiple outcomes, utilizing observa- tional data. The proposed method simultaneously estimates patient-specific weighting of multiple outcomes and the decision-making process, allowing for construction of ITRs with continuous doses. In Chapter 4, we present a new technique for optimizing individualized treatment rules (ITRs) in the context of continuous treatment options. This method leverages the conditional mean of a newly proposed pseudo-outcome to model the counterfactual outcome under the assigned treat- ment rule. By applying kernel regression, the objective function takes an explicit form, allowing it to be eciently solved using conventional optimization algorithms.