The role of artificial intelligence in the diagnosis of rare diseases in general practice using the example of Castleman's disease

Johannes Fluch-Niebuhr, Jean Tori Pantel, Sonja Hermeneit, Norbert Van Rooij

Keywords: Rare diseases Castleman disease Artificial intelligence Large language models Clinical Decision Support Systems

Introduction:

GPs are often the first port of call for patients with non-specific complaints such as fatigue, night sweats or unwanted weight loss. While these symptoms are usually caused by common diseases, they can also be caused by one of around 8,000 rare diseases.
Artificial intelligence (AI) could play a role in the early identification of rare diseases by analyzing clinical data and supporting doctors in differential diagnosis.

Method:

As part of the work, a literature search was carried out on the topic. The extent to which large language models (LLMs) are able to recognize Castlemann's disease on the basis of a simulated patient profile was then investigated.

Results:

In the test series, 7 out of 17 LLMs recognized the disease as a possible differential diagnosis. The diagnostic accuracy varied depending on the prompt optimization and model training.

Conclusions:

The results show that AI-supported systems could support the diagnosis of rare diseases such as Castleman's disease. In some cases, LLMs identified the disease as a differential diagnosis, with accuracy strongly dependent on model architecture and prompt optimization. However, key challenges are the availability of high-quality, standardized data and optimized algorithms for the primary care context. Controlled implementation in primary care could improve the early detection of rare diseases and facilitate targeted referrals to specialists. Further research is needed to evaluate the validity of such models and their integration into existing care structures.

#40