Abstract
This dissertation develops an integrated, data-driven framework for disease management that connects statistical inference, policy evaluation, and sequential decision optimization, applied to small renal mass (SRM) management. SRMs, defined as solid enhancing renal lesions 4 cm or smaller in diameter, represent a rapidly growing clinical challenge. A central management challenge is the competing-risk tradeoff between cancer control and kidney function preservation. Surgical intervention reduces cancer risk but causes irreversible kidney function loss, which in turn elevates cardiovascular mortality. This competing-risk structure makes SRM management an inherently sequential and patient-specific problem that cannot be fully addressed by static, population-level guidelines.
The dissertation makes three primary contributions. These contributions progress from estimating disease dynamics, to evaluating existing care strategies, and finally to optimizing treatment decisions.
The first contribution is an Expectation–Maximization (EM) algorithm for estimating chronic kidney disease (CKD) stage transition probabilities from irregularly observed electronic health record (EHR) data. By treating unobserved intermediate transitions as latent variables, the EM framework estimates CKD progression dynamics from the University of Virginia SRM registry while explicitly accounting for interval censoring and irregular observation intervals. The resulting transition matrices provide model-ready inputs for the downstream decision-analytic models.
The second contribution develops a patient-level microsimulation framework and uses it to systematically compare five major international SRM management guidelines (AUA, NCCN, EAU, ASCO, and CUA). Under common modeling assumptions, the five guidelines produce broadly similar outcomes on both a standard population and a calibrated institutional cohort. However, guideline performance is shown to depend on underlying population characteristics rather than being an intrinsic property of the guidelines themselves. Sensitivity analyses identify the frailty age cutoff and the active-surveillance size limit as particularly influential parameters, while cohort-dependence analyses demonstrate that between-guideline differences increase with tumor burden and preserved renal function. Beyond guideline comparison, the microsimulation framework provides the evaluation platform used to assess optimized treatment policies.
The third contribution formulates SRM management as a finite-horizon Markov Decision Process (MDP) with a seven-dimensional state space capturing tumor size, CKD stage, metastatic status, treatment history, and patient demographics. Three competing mortality channels are jointly modeled, and the action set includes four ordered treatment options: active surveillance (AS), thermal ablation (ABL), partial nephrectomy (PN), and radical nephrectomy (RN). We derive structural properties of the optimal policy, including value function monotonicity in disease severity, existence of a state-dependent intervention threshold, and nondecreasing threshold monotonicity in both patient age and CKD severity. These results provide formal support for treatment strategies that account for patient age and kidney function. The MDP is solved using the Storm probabilistic model checker, representing, to our knowledge, the first application of formal verification tools to a cancer treatment optimization problem. Evaluated within the microsimulation framework, the optimal policy derived from the MDP achieves higher quality-adjusted life-years (QALYs) than the AUA and NCCN guideline strategies, with gains concentrated in patient subgroups where the competing-risk tradeoff is most acute.
Together, these contributions establish an integrated methodological framework that links data-driven estimation of disease dynamics, evaluation of existing guideline-based care, and formal optimization of sequential treatment decisions. The framework demonstrates how statistical inference, simulation, and optimization can be combined to support personalized treatment planning in oncology.
