MAP: evaluation and multi-agent enhancement of large language models for inpatient pathways

Abstract Inpatient pathways require complex clinical decision-making based on comprehensive patient information, yet research on medical LLMs is limited in this area due to the lack of large-scale datasets. Existing medical benchmarks primarily focused on question-answering and examinations, overlooking the multifaceted nature of inpatient decision-making. To address this gap, we developed the IPDS benchmark, comprising 51,274 cases across 9 triage departments, 17 major disease categories, and 16 treatment options. We further proposed the Multi-Agent Inpatient Pathways (MAP) framework, containing three specialized clinical agents: a triage agent for patient admission, a diagnosis agent for diagnostic decision-making, and a treatment agent for care planning. A chief agent guides and promotes these agents to ensure coordination. Experiments demonstrated that MAP achieved superior alignment with operational protocols compared to state-of-the-art LLMs. The MAP sets a foundation for advancing inpatient support systems, offering significant potential for enhancing operational efficiency and resource planning in healthcare facilities.