Date of Award
Fall 2025
Document Type
Open Access Dissertation
Department
Statistics
First Advisor
Jiajia Zhang
Abstract
In precision medicine, developing individualized treatment rules (ITRs) is critical for assigning optimal therapies tailored to patient-specific characteristics. While numerous methods have been proposed for learning ITRs in survival analysis, most focus on binary treatment settings or are limited to a small number of treatment options. In practice, however, patients may face a large menu of potential treatments. As the number of options grows, treatments with similar efficacy may naturally form latent clusters, suggesting that treatment grouping could improve both interpretability and statistical efficiency. These considerations motivate the development of methodologies that can simultaneously handle a large number of treatments and reveal underlying treatment groupings. In this dissertation, we aim to build a comprehensive learning framework for individualized treatment rules in survival settings with optimal treatment grouping. Specifically, we develop parametric, semiparametric, and nonparametric models across three projects.
In the first project, we develop a parametric ITR learning method, AFTSCAF, which extends the Accelerated Failure Time (AFT) framework to settings with many treatments. The key idea is to directly model log-survival time while simultaneously learning treatment groupings through a supervised clustering approach with adaptive fusion (SCAF). This formulation allows AFTSCAF to estimate the optimal ITR that maximizes mean survival while identifying treatments with similar effects. To achieve this, we construct an inverse probability of censoring weighted least squares (WLS) objective and incorporate a flexible fusion penalty that encourages clustering of treatments. The resulting penalized loss function is efficiently optimized via group lasso and accelerated proximal gradient descent.
The second project extends this framework by developing a more flexible Cox proportional hazards (PH)–based model, termed CSCAF. This method builds on the penalized partial likelihood of the Cox PH model and incorporates a supervised clustering approach with adaptive fusion, enabling simultaneous estimation of optimal ITRs and treatment groupings. The optimization problem is efficiently solved using group lasso and accelerated proximal gradient descent. To enhance interpretability, we construct hierarchical clustering plots for both projects that illustrate the treatment grouping process as the tuning parameter varies, and the optimal grouping is selected according to the best tuning parameter.
In the third project, we reformulate ITR learning as a multi-class classification problem and develop a nonparametric method called Survival Group Outcome Weighted Learning (SGROWL) within the GROWL framework. The key idea is to define survival outcomes using the negatives of martingale residuals from a Cox PH regression on baseline covariates (excluding treatment), allowing survival information to be incorporated into outcome-weighted learning. For a fixed treatment grouping, the group-based ITR is estimated using a Reinforced Angle-based Multicategory Support Vector Machine (RAMSVM). To identify the optimal grouping, we implement a greedy search algorithm that iteratively explores partitions and estimates the corresponding group-based ITRs. Finally, across different group numbers, the optimal grouping is selected as the one yielding the largest improvement in mean weighted outcomes compared with the no-grouping scenario.
For all three methods, we conduct extensive simulation studies to evaluate performance, demonstrating accurate estimation of optimal ITRs and reliable recovery of underlying treatment group structures. We further illustrate the practical utility of our framework through a head and neck cancer dataset from the University of North Carolina Hospital, highlighting its applicability to real-world clinical data.
Rights
© 2025, Ziang Liu
Recommended Citation
Liu, Z.(2025). Learning Individualized Treatment Rules With Optimal Treatment Grouping for Survival Outcomes. (Doctoral dissertation). Retrieved from https://scholarcommons.sc.edu/etd/8691