Enhancing sepsis management in the intensive care unit: A multi-agent methodology for decision support

Abstract

Sepsis management in the Intensive Care Unit (ICU) requires prompt and effective decision-making, particularly in the early stages when microbial culture results are unavailable. To enhance decision support in this critical setting, we developed a multi-agent system leveraging specialized agents’ complementary expertise. The system consists of three agents: a Sepsis Management Agent, an Antibiotic Recommendation Agent, and a Guideline Compliance Agent. These agents use retrieval-augmented generation to provide tailored recommendations based on patient information, relevant literature, and clinical guidelines. The system was evaluated using programmatic metrics and human expert assessments. Based on the TruLens framework, the programmatic evaluation showed promising results in terms of answer relevance, context relevance, and groundedness. However, the human expert evaluation revealed a fair level of inter-rater agreement (Cohen's Kappa = 0.21, p-value = 0.02), highlighting the importance of human oversight in assessing the system's outputs. While the multi-agent approach showed promise in enhancing sepsis management decision support, the persistent risk of hallucinations underscores the need for continued refinement and careful deployment of such AI-powered systems in critical care environments. By providing access to the system's documentation and code, we invite further exploration and improvement by the healthcare informatics and biomedical engineering community.