A large language model-driven multidisciplinary AI agent system predicts delirium in emergency critically ill patients

Delirium occurs frequently in emergency departments and is associated with poor outcomes and increased burden. Early delirium risk prediction is crucial for timely prevention and intervention in emergency care, but most existing models focus on intensive care unit (ICU) populations and offer limited interpretability and interactivity. We propose DeLiriuMAgents, a large language model (LLM)-driven multi-agent system for predicting delirium risk in emergency critically ill patients. It simulates multidisciplinary clinical consultation by integrating data-driven, machine learning-based risk prediction; LLM-based virtual specialist reasoning in emergency medicine, neurology, and psychiatry; and medical evidence via retrieval-augmented generation to reach a final decision. In model development, Medical Information Mart for Intensive Care (MIMIC)-IV is used for model derivation and internal validation; a multicenter Peking University (PKU) cohort from two hospitals in China and the eICU Collaborative Research Database (eICU-CRD) cohort are used for external validation. It achieves accuracy/sensitivity/specificity of 0.749/0.762/0.747, 0.731/0.708/0.736, and 0.670/0.708/0.665 on MIMIC-IV, PKU, and eICU-CRD validation sets, respectively. Chart review and clinician evaluation verify the interpretability and usefulness of its reports.