Using Patient Preferences to Estimate Optimal Treatment Strategies for Competing Outcomes Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 21, 2019
  • Butler, Emily
    • Affiliation: Gillings School of Global Public Health, Department of Biostatistics
  • Treatment decisions should be tailored as close as possible to heterogeneous disease populations because diversity permeates through all levels of patient information. This can include genetic makeup, demographic variables, or individual goals as to what qualifies as a successful outcome; hence, patient treatment plans should account for all of these factors. This area of clinical research, coined precision medicine, focuses on combining a multitude of considerations to make treatment decisions as personalized as possible. In this sphere, the statistical contribution involves methodology that most accurately maps patient information to the set of treatment options. While there has been a plethora of research developing personalized treatment plans, they are centralized around creating optimal strategies for only one outcome. There has been limited work done to estimate personalized treatment plans for patients interested in balancing competing outcomes. This work seeks to fill that gap. One way of balancing competing outcomes is to incorporate the patient's preferences regarding these outcomes in the development of a utility function to be used in the estimation procedure. Since it is not possible to directly observe a patient's preference in standardized numerical form, we solve this using a preference elicitation questionnaire in conjunction with item response theory. We derive a posterior estimate of each patient's latent preference information and use it to define a utility function that represents the patient's inherent trade-offs between the outcomes. The optimal treatment choice is that which provides the largest expected utility for each patient, conditional on the patient's prognostic information. This estimation technique is extended to the multi-stage treatment scenario which requires sequential decision making. Estimating the latent preference value now involves the patient's contentment with results from previous stages along with the evolution of the patient's preferences. Once the composite outcome is defined, Q-learning is used to determine which treatment elicits the largest expected utility given the patient's prognostic information, while assuming that the optimal treatment will be chosen in the future. Finally, to make the estimation technique more flexible, we propose a nonparametric approach to both estimating the latent preference and defining the utility function via monotone splines.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Cole, Stephen
  • Davis, Sonia
  • Zeng, Donglin
  • Edwards, Lloyd
  • Laber, Eric
  • Kosorok, Michael
  • Doctor of Philosophy
Degree granting institution
  • University of North Carolina at Chapel Hill Graduate School
Graduation year
  • 2016

This work has no parents.