Decision-Analytic Models for Treatment Optimization in the Presence of Patient Heterogeneity

Abstract

With the ever-increasing complexity in disease etiology, new therapeutics, healthcare service delivery, and clinical guidelines, selecting the appropriate course of treatment for an individual or population can become a great challenge for clinicians and healthcare providers. Applying suboptimal healthcare policies can set a damaging course for infectious disease control on the population level. On the individual level, disease can progress uniquely from patient-to-patient and ignoring a patient’s preference may lead to treatment nonadherence or treatment rejection. In this thesis, we address the need for the development of decision-analytic methods for treatment selection that accounts for diversity in the patient population, uncertainty in patient treatment responses, and patients’ preferences by studying the following problems: 1) the HIV treatment policy selection in HIV-infected children in sub-Saharan Africa initiating treatment at age ≥ 3 years old in the presence of pre-treatment drug resistance; 2) a personalized treatment selection problem for chronic depression where patient’s respond uniquely to treatments whose response level is quantified by their unknown treatment effects; 3) a personalized treatment selection problem with two competing objectives, health outcomes and treatment side effect burden, given the qualitative rankings of sequences of possible patient’s treatment and responses For the first problem, we develop and calibrate a microsimulation model of HIV disease progression and treatment. Using the model, we evaluate alternative antiretroviral treatment strategies using cost-effectiveness analyses. The second problem is formulated as a Markov Decision Process (MDP) where the treatment progression is parametrized by an individual’s unknown treatment effects. We solved for the personalized treatment policies using two heuristic approaches: a model-based approach that can estimate an individual’s treatment effect and a model-free approach using reinforcement learning. Taking into account an individual’s preferences over two objectives, we formulate the last problem as an MDP as well. We developed two search algorithms, exhaustive and heuristic search, to estimate a patient’s preference and provide an optimal treatment plan. This thesis contributes in developing three decision-analytic models to support decisions in testing, monitoring, and treatment selection for two significant healthcare problems, specifically, HIV and chronic depression, and treatment selection incorporating patient’s preference. In addition, our work provides a step towards the design of personalized treatment strategy for patients with chronic diseases in various scenarios.