Adaptive treatment designs for optimal medical care

Effective medical care must flexibly adapt to changing patient characteristics. Optimal treatment strategies in this context should incorporate patient feedback from past treatments to decide future care. In this session, we will explore statistical methods that address the complex dependencies implied by this sequential decision structure. There are multiple ways of efficiently finding the best adaptive treatment strategies. Sequential Multiple Assignment Randomized Trials (SMARTs) randomize participants at sequential decision points to test treatment strategies over time. The data from this trial can then be used to learn optimal adaptive decision rules for future medical interventions. However, in some contexts, it may be necessary or advantageous for researchers to pinpoint and assign the best treatment for each patient during the experimental process. Online reinforcement learning (RL) provides a natural framework for this task, learning from past medical interventions to find the optimal treatment for a given individual in real-time. It achieves this goal by making decisions that efficiently balance exploration of possible solutions and exploitation of current knowledge to assign the best treatments. The specifics of this balance are highly dependent on context. In healthcare, RL methods should be carefully designed to account for complex patient outcomes and minimize patient risk under uncertainty. This session will provide innovative tools for researchers to find optimal decision strategies for medical care over time.