FAIRNESS-AWARE MULTI-AGENT REINFORCEMENT LEARNING FOR EQUITABLE SEPSIS DETECTION ACROSS DEMOGRAPHIC SUBGROUPS

Abstract

Sepsis, a dysregulated host response to infection leading to life-threatening organ dysfunction, is a leading cause of in-hospital mortality in the United States and worldwide, with an estimated one in five deaths globally attributed to the condition.1 The prognosis for septic patients is critically time-dependent, making early and accurate detection paramount for survival.2 While Artificial Intelligence (AI), particularly Reinforcement Learning (RL), has shown considerable promise in optimizing clinical decisions for sepsis treatment, these models often overlook the critical issue of algorithmic fairness.3 The proliferation of AI in healthcare is fraught with the peril of algorithmic bias, which can perpetuate and even exacerbate existing health disparities affecting vulnerable demographic subgroups.4 Existing RL models for sepsis, such as the seminal "AI Clinician," have focused on optimizing a single treatment policy for a general population, without explicitly addressing the challenge of equitable detection performance across diverse racial, gender, and age groups.3