Multi-Agent Clinical Decision Support Systems for ICU Triage: An Agentic AI Framework for Real-Time, Explainable Critical Care Decision-Making

Abstract

Intensive Care Unit (ICU) triage requires clinicians to synthesize heterogeneous, rapidly evolving data streams vital signs, laboratory results, medication histories, and unstructured clinical notes under severe time pressure, where errors can directly translate into preventable mortality. In ICU triage, the combination of heterogeneous, rapidly evolving data streams such as vital signs, lab results, medication history, and unstructured clinical notes must be synthesized under severe time constraints, with the potential for errors to have immediate impact on mortality. Traditional Clinical Decision Support Systems (CDSS) use fixed, rule-based scores (such as APACHE II, SOFA) or a single monolithic Machine Learning (ML) model and are unable to change their behavior to fit the context, provide explanations for their behavior, or gracefully degrade when data is incomplete. The proposed multi-agent clinical decision support architecture for triage in an ICU describes a collection of autonomous agents specialized in monitoring vital signs, interpreting laboratory results, reasoning in the context of the patient, verifying medication safety, data integration, and explaining the process results to the human operator to obtain an overall human-interpretable assessment of the patient’s clinical severity. Based on the recent successes of LLM-based multi-agent systems in emergency and intensive care, the study presents the architecture, orchestration protocol, and evaluation methodology of the system and discusses their development within an ethical AI governance framework that highlights the importance of transparency, fairness, and accountability. This modular, auditable approach to clinical reasoning is more clinically credible and readily scalable to full autonomy in the ICU setting, and three open challenges are highlighted that need to be addressed before it can be used in the real world: latency, safety checks, and regulatory clearance.