Reinforcement Learning for Digital Health Interventions in the Dyadic Setting

Monday, May 5, 2024 – 3:00 pm CT

About the Webinar:

Existing digital health interventions primarily target a single individual. However, many real-world health behaviors occur within social contexts where interventions on connected individuals can improve outcomes of a target individual. An example is the patient-care partner social network, called a dyad. Intervening at the dyadic level requires designing coordinated, personalized interventions that address each member individually and their relationship.

We illustrate this concept through ADAPTS-HCT, an upcoming clinical trial involving adolescents and young adults (AYAs) who have undergone hematopoietic stem cell transplant and their care partners. The trial provides digital interventions to the patient, the caregiver, and their relationship, with the goal of enhancing AYA’s medication adherence. Reinforcement learning (RL) offers a promising approach to personalize intervention delivery dynamically. A key methodological challenge arises because certain intervention components (e.g., those targeting caregivers) do not directly impact adherence, resulting in distal treatment effects. To mitigate this issue, we present a hierarchical multi-agent RL framework for sequential decision-making within social networks. Our framework efficiently leverages the structure of dyadic networks, enabling fast learning in these complex settings.

About the Presenter:

Ziping is a Postdoctoral Fellow in Statistics at Harvard University, working with Professor Susan Murphy. He earned his PhD in Statistics from the University of Michigan under the guidance of Professor Ambuj Tewari. His research focuses on sequential decision-making methods, such as Reinforcement Learning, with applications to digital intervention delivery in mobile health. He collaborates closely with domain experts to design and implement Reinforcement Learning agent in online mobile health clinical trials. More about Ziping Xu.