A new research paper explores the challenges of selective prediction in rare-disease diagnosis using AI. The study found that even advanced open-weight LLMs struggle with ultra-rare diseases, achieving low recall rates. The paper also highlights that relying solely on the top score for diagnostic decisions is unreliable, and a more nuanced approach using confidence signals like the top-two margin is necessary for improved accuracy. The research suggests that current methods are insufficient for reliably identifying rare diseases and that further model development is needed. AI
IMPACT Highlights limitations of current LLMs in specialized diagnostic tasks, indicating a need for more robust models for rare disease identification.
RANK_REASON Research paper published on arXiv detailing limitations of AI in rare-disease diagnosis. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
