A Non-Parametric Framework for Value-Based Individualized Treatment Rule Estimation

Abstract

The objective of precision medicine is to match patient and treatment such that optimal outcomes for the patient are achieved. Precision medicine represents a departure from the 'one size fits all' paradigm in which a single treatment recommendation is made for all patients presenting with a given disease or set of symptoms. While this previous paradigm may lead to positive outcomes on average, heterogeneity of treatment effects for individual patients may lead to varied results for certain patients. In support of precision medicine, many statistical methods appear in the literature, often denoted as individualized treatment rules (ITRs). While optimal performance of an ITR is of great importance, interpretability of the treatment rule, an understanding of 'why' a treatment path is being recommended, is also of concern to patients. Despite a rich literature concerning ITR estimation, there is a lack of methods focusing on generating interpretable ITRs. Additionally, many methods rely on regression techniques that may be subject to model misspecification. To address these gaps in the literature, the methods presented in this document construct interpretable ITRs using decision trees with a direct optimization purity measure proposed for each type of outcome. An interpretable ITR framework is developed for a typical univariate, continuous outcome with a binary treatment, extended to a risk constrained optimizer, and additionally modified to incorporate time-to-event data with time varying covariates. The proposed ITR estimation techniques are organized into three projects. The first project considers a univariate, continuous, efficacious outcome to be optimized against a binary treatment. A novel splitting function is proposed to maximize expected efficacy, with inverse probability weighted (IPW) and augmented IPW estimators of efficacy provided. Performance of the proposed methods is assessed via a simulation study, and a data analysis is performed using observations from both randomized, controlled trial (RCT) and observational electronic medical records (EMR) sources. The second project optimizes an efficacious outcome under a risk constraint by introducing a tuning parameter that penalizes rules with excess expected risk relative to expected efficacy. Again, a simulation study and data analysis are provided to illustrate performance. The third project considers time-to-event endpoints, and provides treatment rules using a restricted mean survival time (RMST) based purity measure. Inclusion of time-varying covariates is accomplished through partitioning individual patients into pseudo-observations and evaluating the cumulative hazard. A simulation study is provided for the RMST method, and a data analysis is presented to demonstrate the sufficiency of the proposed methods. Collectively, this work constitutes a set of interpretable, non-parametric ITR estimation procedures with application to varying clinical outcomes arising from RCT or EMR data sources.