IMPROVING RISK PREDICTION AND TREATMENT DECISIONS USING AUXILIARY INFORMATION

Precision medicine hinges on two essential components: risk prediction and individualized treatment rules (ITR). Risk prediction identifies individuals at higher risk for adverse health outcomes, while ITRs customize treatments based on individual characteristics. Incorporating auxiliary information can improve both processes, leading to more informed clinical decisions. The dissertation addresses critical challenges in these domains through three interconnected projects. The first project develops a novel framework for event time prediction using auxiliary time-dependent covariates. Using the support vector machine and the kernel trick, this method predicts associated counting processes while accommodating nonlinear relationships in both time and covariate domains. A chain algorithm is employed to predict future events, and theoretical analysis establishes its equivalence with comparing time-varying hazard rates among at-risk subjects. Simulations and a case study on Huntington’s disease highlight superior predictive accuracy compared to state-of-the-art methods in machine learning, deep learning, and statistical models. The second project focuses on the evaluation of single-stage ITR derived from randomized clinical trials (RCTs) when applied to broader target populations characterized by auxiliary real-world data (RWD). The distributional differences between RCTs and RWD can obscure the true benefit of these ITRs. To address this, we propose an unbiased estimator with inferential properties and introduce two novel composite features. These features, in replacement of the potentially high-dimensional original features, not only yield an unbiased evaluation of ITR but also minimize the variability of its estimated benefits. Simulations illustrate the method’s robustness in small samples, and the composite features, if learned apriori, show potential for informing future trial designs. The third project introduces a framework for multistage dynamic treatment regimes (DTRs) that addresses biases and lack of generalizability in rules estimated from sequentially multiple-assignment randomized trials (SMART) due to covariate distributional discrepancies. By integrating SMART data with RWD and applying importance weighting, we develop interpretable DTRs consistent with the optimal linear DTRs in the target population. Simulation studies and an application to type II diabetes treatment highlight the superior performance of the approach over existing methods. Together, these projects advance risk prediction and treatment personalization, laying the foundations for more effective precision medicine.