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

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Neurology and Stroke Management

Background:

• Acute ischemic stroke (AIS) treatment selection is complex, requiring rapid integration of diverse clinical data.
• Current decision-making processes are expertise-dependent and prone to critical errors.

Purpose of the Study:

• To develop and validate a structured multiagent large language model (LLM) framework for AIS decision support.
• To assess the LLM framework's accuracy, safety, auditability, and impact on physician decision-making, particularly for less experienced clinicians.

Main Methods:

• A multiagent LLM framework was developed using structured outputs and guideline-based reasoning for treatment recommendations and TOAST classification.
• The framework was evaluated on retrospective, prospective, and literature-derived AIS cases.
• Physician decision-making with and without LLM support was assessed in a prospective study.

Main Results:

• The LLM framework consistently improved treatment recommendation accuracy across multiple models and case types (e.g., 0.546-0.737 to 0.687-0.851 in group A).
• LLM support increased physician treatment decision accuracy from 73.1% to 88.6%, with the most significant gains among junior and nonspecialist physicians.
• The framework improved clinical safety scores and reduced hallucination and omission rates.

Conclusions:

• A structured multiagent LLM framework enhances AIS treatment recommendations and classification, improving accuracy and safety.
• The framework shows potential to reduce expertise-related decision accuracy gaps among physicians.
• Further multicenter studies are needed to evaluate workflow and clinical outcome impacts.