Scholarlink

Background <p>Rare diseases such as Castleman disease are difficult to detect in primary care, as they often present nonspecific symptoms like fatigue, night sweats, or weight loss. Early suspicion can improve prognosis. Artificial intelligence (AI), especially large language models (LLMs), may support the differential diagnosis by systematically analyzing clinical information.</p> Objective <p>This exploratory pilot study aimed to investigate whether freely available LLMs can identify rare diseases such as Castleman disease as differential diagnoses in the primary care setting and suggest appropriate referral targets.</p> Methods <p>A&#xa0;standardized case vignette was developed based on the diagnostic criteria of the German Society for Hematology and Medical Oncology (DGHO) for iMCD (idiopathic multicentric Castleman disease) and validated in interviews with three affected individuals. Nine general practitioners (GPs) were asked about their diagnostic considerations and referral decisions. The same vignette was then submitted to 17&#xa0;freely available LLMs using standardized prompts. The generated diagnoses and referral recommendations were evaluated qualitatively and using quantitative descriptive methods. A&#xa0;narrative literature review provided context.</p> Results <p>The suspected diagnoses of GPs were dominated by general hematological and oncological conditions (60%), but no specific suspected diagnoses were mentioned. The LLMs suggested specific oncological or hematological differential diagnoses in 53.5% of the generated differential diagnoses, including iMCD (11.6%). A&#xa0;referral to hematology/oncology was most frequently recommended (LLMs 40.9%; GPs 65.2%).</p> Conclusion <p>The results provide insights into how generative language models can raise awareness of rare differential diagnoses and support decision-making in primary care. However, their use requires medical contextualization. Due to the small number of cases and heterogeneous systems, the results should be interpreted as exploratory. Future studies should include larger samples, multiple case vignettes, and standardized evaluation protocols to systematically test the practical potential of AI applications in primary care.</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

Früherkennung seltener Erkrankungen mit künstlicher Intelligenz

  • 期刊论文

  • 发表日期：

    2026年2月4日

查看全文
  • Jean Tori Pantel,
  • Kai Hübel,
  • Johannes Fluch-Niebuhr,
  • Marcus Hentrich,
  • Paul Liedgens,
  • Sonja Hermeneit,
  • Thomas Lipp,
  • Dorit Maoz,
  • Jan Kirchhoff,
  • Martin Mücke,
  • Laura Dalhaus,
  • Ulrike Rott,
  • Norbert van Rooij

摘要

Background

Rare diseases such as Castleman disease are difficult to detect in primary care, as they often present nonspecific symptoms like fatigue, night sweats, or weight loss. Early suspicion can improve prognosis. Artificial intelligence (AI), especially large language models (LLMs), may support the differential diagnosis by systematically analyzing clinical information.

Objective

This exploratory pilot study aimed to investigate whether freely available LLMs can identify rare diseases such as Castleman disease as differential diagnoses in the primary care setting and suggest appropriate referral targets.

Methods

A standardized case vignette was developed based on the diagnostic criteria of the German Society for Hematology and Medical Oncology (DGHO) for iMCD (idiopathic multicentric Castleman disease) and validated in interviews with three affected individuals. Nine general practitioners (GPs) were asked about their diagnostic considerations and referral decisions. The same vignette was then submitted to 17 freely available LLMs using standardized prompts. The generated diagnoses and referral recommendations were evaluated qualitatively and using quantitative descriptive methods. A narrative literature review provided context.

Results

The suspected diagnoses of GPs were dominated by general hematological and oncological conditions (60%), but no specific suspected diagnoses were mentioned. The LLMs suggested specific oncological or hematological differential diagnoses in 53.5% of the generated differential diagnoses, including iMCD (11.6%). A referral to hematology/oncology was most frequently recommended (LLMs 40.9%; GPs 65.2%).

Conclusion

The results provide insights into how generative language models can raise awareness of rare differential diagnoses and support decision-making in primary care. However, their use requires medical contextualization. Due to the small number of cases and heterogeneous systems, the results should be interpreted as exploratory. Future studies should include larger samples, multiple case vignettes, and standardized evaluation protocols to systematically test the practical potential of AI applications in primary care.