A0749
Title: Preference-integrated dynamic treatment regimes: Methodology and application to SMARTs
Authors: Yating Zou - University of North Carolina at Chapel Hill (United States) [presenting]
Michael Kosorok - University of North Carolina at Chapel Hill (United States)
Lesile Wilson - University of California San Francisco (United States)
Joshua Zitovsky - Amazon (United States)
Abstract: In modern healthcare, it is often useful for treatment decisions to balance multiple outcomes according to patient preferences. For example, efficacy and side effects. The purpose is to introduce latent utility Q-learning (LUQ-Learning), a Q-learning algorithm augmented by latent variable modeling, that embeds random individual preferences into the optimality criterion. LUQ-Learning supports a finite number of decision points and finite dimensions of outcomes with asymptotic guarantees under realistic assumptions. In simulations, LUQ-Learning consistently outperforms alternatives. The method was applied to BEST, a SMART study on chronic low back pain, highlighting its practical utility in precision medicine.
