Title: Assessing how well RL algorithms personalize in mobile health
Authors: Zeyang Jia - Harvard University (United States)
Peng Liao - Harvard University (United States)
Kelly Zhang - Harvard University (United States)
Susan Murphy - Harvard University (United States)
Abstract: Reinforcement learning (RL) algorithms are increasingly used in mobile health applications. By adaptively personalizing the delivery of interventions to users, RL algorithms are designed to learn which intervention strategies are most effective in different circumstances for improving users' health outcomes. Therefore, after using an RL algorithm in a mobile health study, it is critical to assess how well the RL algorithm is at personalizing interventions to users: Does it learn over time and achieve better outcomes for users than a non-RL algorithm does? To answer this question, we formally define the meaning of personalization in mobile health and propose statistical methods for investigating whether an algorithm personalizes. We also apply our method to data from a mobile health clinical trial and assess the personalization of the algorithm in the study.