Grounded large language models for diagnostic prediction in real-world emergency department settings.

Objective: To evaluate predictive diagnostic performance of open- and closed-source large language models (LLMs) in emergency medicine, addressing the urgent need for innovative clinical decision support tools amid rising patient volumes and staffing shortages. Materials and Methods: We generated 2370 AI-driven diagnostic predictions (Top-5 diagnoses from each of 6 model pipelines per patient), using data from 79 real-world emergency department cases collected consecutively during a 24-hour peak influx period at a tertiary care center. Pipelines combined open- and closed-source embedding models (text-embedding-ada-002, MXBAI) with foundational models (GPT-4, Llama3, and Qwen2) grounded via retrieval-augmented generation using emergency medicine textbooks. Models' predictions were assessed against reference diagnoses established by expert consensus. Results: All pipelines achieved comparable diagnostic match rates (62.03%-72.15%). Diagnostic performance was significantly influenced by case characteristics: match rates were notably higher for specific versus unspecific diagnoses (85.53% vs 31.41%, Discussion: Diagnostic accuracy primarily depended on case characteristics rather than the choice of model pipeline, highlighting fundamental AI alignment challenges in clinical reasoning. Low performance in unspecific diagnoses underscores inherent complexities in clinical definitions rather than technological shortcomings alone. Conclusion: Open-source LLM pipelines provide enhanced sourcing capabilities, crucial for transparent clinical decision-making and interpretability. Further research should expand knowledge bases to include hospital guidelines and regional epidemiology, while exploring on-premises solutions to better align with privacy regulations and clinical integration.