Background Emergency departments face increasing pressures from staff shortages, patient surges, and administrative burdens. While large language models (LLMs) show promise in clinical support, their deployment in emergency medicine presents technical and regulatory challenges. Previous studies often relied on simplistic evaluations using public datasets, overlooking real-world complexities and data privacy concerns.

Methods At a tertiary emergency department, we retrieved 79 consecutive cases during a peak 24-hour period constituting a siloed dataset. We evaluated six pipelines combining open- and closed-source embedding models (text-embedding- ada-002 and MXBAI) with foundational models (GPT-4, Llama3, and Qwen2), grounded through retrieval-augmented generation with emergency medicine textbooks. The models’ top-five diagnostic predictions on early clinical data were compared against reference diagnoses established through expert consensus based on complete clinical data. Outcomes included diagnostic inclusion rate, ranking performance, and citation sourcing capabilities.

Results All pipelines showed comparable diagnostic inclusion rates (62.03-72.15%) without significant differences in pairwise comparisons. Case characteristics, rather than model combinations, significantly influenced predictive diagnostic performance. Cases with specific diagnoses were significantly more diagnosed versus unspecific ones (85.53% vs. 31.41%, p<0.001), as did surgical versus medical cases (79.49% vs. 56.25%, p<0.001). Open-source foundational models demonstrated superior sourcing capabilities compared to GPT-4-based combinations (OR: 33.92 to ∞, p<1.4e-12), with MBXAI/Qwen2 achieving perfect sourcing.

Conclusion Open and closed-source LLMs showed promising and comparable predictive diagnostic performance in a real-world emergency setting when evaluated on siloed data. Case characteristics emerged as the primary determinant of performance, suggesting that current limitations reflect AI alignment fundamental challenges in medical reasoning rather than model-specific constraints. Open-source models’ demonstrated superior sourcing capabilities—a critical advantage for interpretability. Continued research exploring larger-scale, multi-centric efforts, including real-time applications and human-computer interactions, as well as real- world clinical benchmarking and sourcing verification, will be key to delineating the full potential of grounded LLM-driven diagnostic assistance in emergency medicine.

The authors have declared no competing interest.

This study did not receive any funding.

I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.

Yes

The details of the IRB/oversight body that provided approval or exemption for the research described are given below:

Ethics committee of CHU UCL Namur gave ethical approval for this work (187/2023)

I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.

Yes

I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).

Yes

I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.

Yes

We have elected not to make our original patient-level dataset publicly available. As this dataset originates from real-world emergency department cases, it contains sensitive clinical information despite de-identification. We are particularly concerned that sharing these data may lead to unintended contamination and dissemination among external models used for training purposes. Such dissemination could compromise the integrity of the dataset and preclude its future use in further research. Consequently, we have decided to restrict access to the data to preserve its unique value and to ensure that its subsequent application in research remains controlled and ethical.