Related Experiment Video
Updated: May 27, 2025
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Evaluating large language model performance to support the diagnosis and management of patients with primary immune
Nicholas L Rider1, Yingya Li2, Aaron T Chin3
1Department of Health Systems and Implementation Science, Virginia Tech Carilion School of Medicine, Roanoke, Va; Section of Allergy and Immunology, Department of Medicine, Carilion Clinic, Roanoke, Va.
Background:
Generative artificial intelligence (GAI) is transforming health care in a variety of ways; however, the present utility of GAI for supporting clinicians who treat rare disease such as primary immune disorders (PIs) is not well studied. We evaluated the ability of 6 state-of-the-art large language models (LLMs) for providing clinical guidance about PIs.
Objective:
To quantitatively and qualitatively measure the utility of current, open-source LLMs for diagnosing and providing helpful clinical decision support about PIs.
Methods:
Five expert clinical immunologists each provided 5 real-world, anonymized PI case vignettes via multi-turn prompting to 6 LLMs (OpenAI GPT-4o, Llama-3.1-8B-Instruct, Llama-3.1-70B-Instruct, Mistral-7B-Instruct-v0.3, Mistral-Large-Instruct-2407, Mixtral-8x7B-Instruct-v0.1). We assessed the diagnostic accuracy of the LLMs and the quality of clinical reasoning using the Revised-IDEA (R-IDEA) score. Qualitative LLM assessment was made by immunologist narratives.
Results:
Performance accuracy (>88%) and R-IDEA scores (≥8) were superior for 3 models (GPT-4o, Llama-3.1-70B-Instruct, Mistral-Large-Instruct-2407), with GPT-4o achieving the highest diagnostic accuracy (96.2%). Conversely, the remaining 3 models fell below acceptable accuracy rates near 60% or lower and had poor R-IDEA scores (≤0.55), with Mistral-7B-Instruct-v0.3 attaining the worst diagnostic accuracy (42.3%). Compared with the 3 best-performing LLMs, the 3 worst-performing LLMs had a substantially lower median R-IDEA score (P < .001). Interclass correlation coefficient for R-IDEA score assignments varied substantially by LLM, ranging from good to poor agreement, and did not appear to correlate with either diagnostic accuracy or median R-IDEA score. Qualitatively, immunologists identified several themes (eg, correctness, differential diagnosis appropriateness, relative conciseness of explanations) of relevance to PIs.
Conclusions:
LLM can support diagnosis and management of PIs; however, further tuning is needed to optimize LLMs for best practice recommendations.
