A Survey for Large Language Models in Biomedicine

LLMs show promise in zero-shot learning across diverse biomedical tasks — The survey analyzed 137 key studies demonstrating that large language models can perform diagnostic assistance, drug discovery, and personalized medicine applications without task-specific training, though performance varies by application.

Fine-tuning and adaptation are necessary for specialized biomedical contexts — While zero-shot capabilities are valuable, the research indicates that customizing LLMs through fine-tuning methods (for both single and multi-modal models) is essential for high-stakes applications like medical question answering and biomedical literature analysis where baseline performance is insufficient.

Data privacy and interpretability remain critical barriers to clinical deployment — The survey identifies data privacy concerns and limited model interpretability as major challenges in biomedicine, where understanding why a model makes a recommendation is often as important as the recommendation itself, particularly given the sensitive nature of patient information.

Dataset quality and ethical concerns require urgent attention — The research highlights that issues with dataset quality and ethical considerations are significant obstacles, reflecting the high stakes of deploying AI in healthcare where errors can directly impact patient outcomes and where bias in training data can perpetuate healthcare disparities.

Comprehensive real-world evaluation is needed before clinical implementation — Based on analysis of 484 publications, the survey emphasizes that practical implications in real-world biomedical contexts must be thoroughly examined, suggesting current research often focuses on isolated applications rather than integration into actual clinical workflows.