Managing uncertainty in sequential medical decision making
- Many models currently used to design and analyze health policies ignore uncertainty in patient outcomes, assume homogeneous patient response to interventions, and do not allow for sequential decision making. However, patient response to treatment is often highly variable; patient outcomes depend on various patient characteristics that can evolve stochastically over time; and decision makers need to respond to new states of the patient as they occur. In this dissertation, we apply stochastic optimization methods to design treatment policies that are adaptive to key patient characteristics in two health settings: treating HIV patients while considering potential long-term cardiovascular side effects and treating multiple sclerosis patients while adapting to their response to treatment. Antiretroviral therapy (ART) for HIV may increase the risk of cardiovascular morbidity and mortality, but delaying ART may diminish immunological benefits. The timing of ART initiation that balances these risks and benefits and yields maximum quality-adjusted life expectancy (QALE) is currently unknown. In Chapter 2, we develop a mathematical model to identify the timing of ART initiation that optimizes patient outcomes as a function of patient CD4 count, age, cardiac mortality risk, gender and personal preferences. Our goal is to find the conditions that maximize patient QALE. We find that, under the assumption that ART confers disease progression and mortality benefits at any CD4 count, immediate treatment initiation yields the greatest remaining QALE for young patients under most circumstances. However, delaying treatment initiation is preferable for older patients with high CD4 counts. The exact timing of ART initiation depends on the magnitude of benefit from ART at high CD4 counts, the magnitude of increases in cardiac risk, and patients' preferences. If ART reduces HIV progression at high CD4 counts, immediate ART is preferable for most newly infected individuals who are less than 35 years old even if ART doubles age- and gender-specific cardiac risk. In Chapter 3, we consider a class of chronic diseases where available treatments are effective only for a subgroup of patients, and biomarkers that accurately assess the responsiveness of an individual patient do not exist. In these settings, information regarding the response type of a patient can only be generated by experimentation ‒ subjecting the patient to a variety of treatments. Physicians then learn about patient response through self-reported patient evaluations, as well as from the occurrence or nonoccurrence of negative health events such as disease flare-ups. The timing of these events also provides substantial information, which should be taken into account when determining optimal personalized treatments. We introduce a continuous-time, two-armed bandit framework that balances the trade-off between exploring alternative treatments and exploiting available information. Unlike most multi-armed bandit models that learn only from observed rewards, our model also incorporates information regarding the frequency of health events, and can be analyzed in closed form to derive guidelines for treatment policies. We illustrate the effectiveness of our methodology by developing an adaptive policy to treat multiple sclerosis, a chronic autoimmune disease. We compare the performance of our policy to that of a standard, non-adaptive treatment policy and show that, by identifying non-responders earlier, our approach leads to improvements in QALE, as well as significant cost savings. Beyond multiple sclerosis, dynamic learning models that incorporate the timing of events may have applications in a broader medical decision making context: for example, as a means to design treatment policies for chronic diseases such as depression, rheumatoid arthritis, celiac disease or Crohn's disease. We conclude with a discussion of the work and directions for future research in Chapter 4.
|Type of resource
|electronic; electronic resource; remote
|1 online resource.
|Negoescu, Diana Maria
|Stanford University, Department of Management Science and Engineering.
|Brandeau, Margaret L
|Brandeau, Margaret L
|Statement of responsibility
|Diana Maria Negoescu.
|Submitted to the Department of Management Science and Engineering.
|Thesis (Ph.D.)--Stanford University, 2014.
- © 2014 by Diana Maria Negoescu
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...