A Multiagent Large Language Model Framework for Emergency Treatment Recommendation in Acute Ischemic Stroke: Development and Validation Study.

Background: Acute ischemic stroke (AIS) treatment selection requires rapid, guideline-concordant integration of clinical, imaging, and laboratory data, including therapeutic windows, contraindications, stroke severity, and imaging eligibility. This process is complex, expertise-dependent, and vulnerable to safety-critical errors. Objective: This study aimed to develop and validate a structured multiagent large language model (LLM) framework for AIS decision support using real-world cases and to assess its accuracy, safety, auditability, and impact on physician decision-making, particularly among junior physicians and nonspecialists. Methods: We developed a multiagent LLM framework that used structured outputs and guideline-based reasoning to generate treatment recommendations (intravenous thrombolysis, endovascular thrombectomy, standard medical therapy, or non-AIS, or nonstroke) and Trial of ORG 10172 in Acute Stroke Treatment (TOAST) classification. The framework was evaluated using multicenter retrospective real-world cases from 2 hospitals collected between January 2018 and March 2025, prospective cases from February to May 2025, and literature-derived challenging cases from PubMed between January 2024 and January 2025. Performance was assessed against clinical reference standards. Safety was assessed using omission and hallucination event rates, instruction adherence, and 5-point clinical safety ratings. In a prospective physician study, physicians with different seniority and specialty backgrounds made AIS treatment and TOAST classification decisions with and without LLM support. Physician-case decision-level outcomes were analyzed using a binomial generalized linear mixed-effects model accounting for physician and case effects. Results: The final analysis included 1055 group A cases, 721 group B cases, 144 literature-derived group C cases, and 161 prospectively collected group D cases. Across representative Baichuan, Qwen, DeepSeek, and GPT models, the multiagent framework consistently improved treatment recommendation accuracy. Model-level accuracy ranges increased from 0.546-0.737 to 0.687-0.851 in group A, from 0.587-0.698 to 0.671-0.813 in group B, and from 0.507-0.646 to 0.667-0.750 in group C. TOAST classification improved overall, with cohort-level variation. Across evaluated models, the multiagent framework increased the mean clinical safety score from 3.70 to 4.01 and reduced mean hallucination and omission rates from 33.6% to 20.6% and from 38.5% to 24.5%, respectively. In the prospective physician study, LLM support increased treatment decision accuracy from 73.1% to 88.6% (odds ratio 2.86, 95% CI, 2.27-3.60; P<.001). Accuracy gains were largest among junior and nonspecialist physicians, including junior specialists (0.667 to 0.833), junior nonspecialists (0.600 to 0.846), and senior nonspecialists (0.667 to 0.850). TOAST classification performance also improved (odds ratio 3.63, 95% CI 2.85-4.64; P<.001). Conclusions: A structured multiagent framework improved LLM performance with average improvements of 18.9% in AIS treatment recommendation and TOAST classification, while producing more structured, auditable outputs with higher safety ratings. It was associated with higher physician decision accuracy, with larger gains among less-experienced physicians, suggesting the potential to narrow expertise-related decision accuracy gaps. Prospective multicenter studies are needed to assess effects on workflow and clinical outcomes.