Reproduce “Large Language Models for Disease Diagnosis: A Scoping Review” on OpenResearch

About this paper

Shuang Zhou, Zidu Xu, Mian Zhang, Chunpu Xu, Yawen Guo, Zaifu Zhan, Yi Fang, Sirui Ding and 9 more

A comprehensive scoping review maps the current landscape of Large Language Model (LLM) applications in disease diagnosis, systematically analyzing 398 studies to identify prevalent techniques, data modalities, and evaluation methodologies across various clinical specialties. It identifies significant research gaps and provides a blueprint for future development in this domain.

Key insights

• Prompting, particularly Chain-of-Thought, is widely utilized for initial explorations and differential diagnosis, while Retrieval-Augmented Generation (RAG) effectively enhances accuracy and mitigates hallucinations by integrating external knowledge.
• Fine-tuning (SFT, RLHF, PEFT) customizes LLMs for specific diagnostic tasks, with pre-training aiming to build comprehensive medical knowledge, though it is resource-intensive.
• Current research heavily relies on automated evaluation metrics but often overlooks critical aspects such as explainability, trustworthiness, privacy, fairness, and the integration of diverse clinical data for robust models.

Summary by alphaXiv.