University of Minnesota Finds Only 12 Studies on LLMs for Rare Diseases
A scoping review finds just 12 studies applying LLMs to rare disease patient communication, revealing a field dominated by ChatGPT question-answering with critical gaps.
ProTakeaways
- A scoping review found only 12 studies on LLMs for rare disease patient education, versus 201 in general patient education.
- ChatGPT was used in all 12 studies; other models appear only once or twice each.
- Reported accuracy on rare-disease questions ranges roughly 70 to 100 percent, but methods are narrow.
- Only one study used retrieval from a curated knowledge base (Orphanet) or domain fine-tuning with LoRA.
- Nine studies were English-only, a critical gap for diseases concentrated in non-English-speaking populations.
- Patients themselves were almost never included as evaluators; readability and empathy were rarely measured.
Rare diseases sit in a strange corner of medicine: individually uncommon, collectively enormous. A new scoping review from a University of Minnesota and Université Paris Cité team argues that large language models could meaningfully help patients and families navigating them, but the research to prove it barely exists yet.
After searching seven major databases across four years, the authors found only 12 studies worldwide on LLM-based rare disease patient education and communication. For context, a 2024 scoping review identified 201 studies using LLMs for patient education in general medical domains, and a 2025 focused review of LLM-generated patient education materials identified 69. Rare disease, despite affecting hundreds of millions of people, is essentially uncharted territory.
A communication nightmare by design
More than 7,000 rare diseases collectively affect an estimated 300 million people worldwide. About 80% have a genetic origin, and more than 90% lack approved treatments. Patients cycle through what clinicians call a diagnostic odyssey, meeting doctors who have never seen their condition, reading scattered literature, and making decisions under deep uncertainty.
LLMs look built for exactly this kind of problem, synthesizing sparse, fragmented information into conversational answers that adapt to the reader. General-purpose models trained on web data are also inherently weak on rare conditions, because the underlying evidence is thin, outdated, or locked inside specialist journals.