Abstract

As healthcare systems move to deploy reinforcement learning (RL) for dynamic resource allocation, significant concerns have been raised about the impact of these tools on health services access and equity. While RL may improve service personalization, black-box optimization based on historical data can inadvertently lead to withholding needed treatment from vulnerable subgroups. We investigated the impact on fairness in treatment allocation using clinical data from a study that tested an RL-supported intervention for managing chronic pain and opioid misuse risk. In this talk, I show how to (1) audit standard RL policies for allocation bias across patient subgroups, and (2) validate a deployable methodology for correcting identified disparities without compromising clinical outcomes. We show that standard methodologies seeking to achieve ”fairness through unawareness” are insufficient. Our framework evaluates and patches these vulnerabilities prior to deployment, satisfying a key requirement for the ethical scaling of AI in healthcare.