Scholarlink

Objectives <p>This study aimed to compare the diagnostic accuracy and citation integrity of four large language models (LLMs) including one general (ChatGPT-4) and three intended for clinical and research use (OpenEvidence, Perplexity, and Pathway), using standardized otolaryngology clinical vignettes.</p> Methods <p>One hundred validated otolaryngology clinical vignettes were presented to each LLM with a prompt requesting both a diagnosis and supporting citations. Diagnostic accuracy was determined against reference answers, and errors were categorized as logical, informational, or explicit. Citation number, source type, hallucination rate, and journal CiteScores were also compared.</p> Results <p>All models demonstrated high diagnostic accuracy (82.0%–91.0%), with ChatGPT-4 achieving the highest numerical accuracy (91.0%), though differences between models were not statistically significant (p = 0.057). Logical errors were most frequent across all models. OpenEvidence and Perplexity generated the most citations per response, while ChatGPT-4 produced the fewest and had the highest hallucination rate (23.0%). Source preferences varied, with OpenEvidence and Pathway favoring narrative reviews and Perplexity favoring government/public health websites. OpenEvidence had the highest mean CiteScore for journal citations.</p> Conclusion <p>This is the first study to assess both diagnostic accuracy and citation integrity of LLMs in otolaryngology. While ChatGPT-4 was most accurate, it had the highest rate of citation hallucinations, suggesting a trade-off between accuracy and source reliability. OpenEvidence, though slightly less accurate, provided more consistent and verifiable references, demonstrating the ability to prioritize citation integrity alongside diagnostic performance for clinical integration.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Diagnostic accuracy and citation integrity of four large language models on otolaryngology vignettes

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • William Pennington-FitzGerald,
  • Akshay Warrier,
  • Sally Durant,
  • Ibraheem Sharaf,
  • Francesca Carlino,
  • Sai Dheeraj Pamula,
  • Jean Anderson Eloy,
  • Jessica Levi

摘要

Objectives

This study aimed to compare the diagnostic accuracy and citation integrity of four large language models (LLMs) including one general (ChatGPT-4) and three intended for clinical and research use (OpenEvidence, Perplexity, and Pathway), using standardized otolaryngology clinical vignettes.

Methods

One hundred validated otolaryngology clinical vignettes were presented to each LLM with a prompt requesting both a diagnosis and supporting citations. Diagnostic accuracy was determined against reference answers, and errors were categorized as logical, informational, or explicit. Citation number, source type, hallucination rate, and journal CiteScores were also compared.

Results

All models demonstrated high diagnostic accuracy (82.0%–91.0%), with ChatGPT-4 achieving the highest numerical accuracy (91.0%), though differences between models were not statistically significant (p = 0.057). Logical errors were most frequent across all models. OpenEvidence and Perplexity generated the most citations per response, while ChatGPT-4 produced the fewest and had the highest hallucination rate (23.0%). Source preferences varied, with OpenEvidence and Pathway favoring narrative reviews and Perplexity favoring government/public health websites. OpenEvidence had the highest mean CiteScore for journal citations.

Conclusion

This is the first study to assess both diagnostic accuracy and citation integrity of LLMs in otolaryngology. While ChatGPT-4 was most accurate, it had the highest rate of citation hallucinations, suggesting a trade-off between accuracy and source reliability. OpenEvidence, though slightly less accurate, provided more consistent and verifiable references, demonstrating the ability to prioritize citation integrity alongside diagnostic performance for clinical integration.