Résumé :

Abstract: In this talk, we will look at our ongoing research addressing the challenges of providing effective chronic pain care through a personalized, equitable, and adaptive approach using reinforcement learning (RL). Chronic pain significantly impacts the quality of life for millions globally, with many individuals lacking access to evidence-based treatments. RL offers promising potential to tailor pain management interventions to individual patient needs while optimizing the use of limited clinical resources. However, concerns persist among clinicians, patients, and healthcare policymakers that RL-based solutions could (inadvertently) exacerbate disparities related to characteristics such as gender or race. Currently, our research focuses on advancing gender fairness in personalized pain care recommendations through a real-world application of RL.
In this context, gender fairness entails minimizing or eliminating disparities in the utility achieved by different gender subpopulations. We investigate how the selection of patient features used in decision-making influences gender fairness. Our proposed RL framework offers two key contributions: (i) the ability to adaptively learn and select features that optimize both utility and fairness simultaneously, and (ii) the capability to accelerate the feature selection process, thereby improving pain care recommendations from the outset by leveraging clinicians' domain expertise. Additionally, we will discuss critical future directions from both machine learning and clinical perspectives.