Medical Entity Linking in Low-Resource Settings with Fine-Tuning-Free LLMs.

Area of Science:

• Biomedical Natural Language Processing
• Artificial Intelligence
• Medical Informatics

Background:

• Medical entity linking aligns text mentions with standardized ontology concepts.
• Existing methods often require extensive data and computational resources.
• Supervised models and domain-specific embeddings are common but resource-intensive.

Purpose of the Study:

• Investigate Large Language Models (LLMs) for medical entity linking.
• Enhance candidate generation and disambiguation using LLMs via synonym expansion and in-context learning.
• Evaluate LLM-based approach against traditional string-matching and supervised methods.

Main Methods:

• Combined string matching with LLMs using in-context learning, avoiding fine-tuning.
• Expanded mention spans with LLM-generated synonyms for candidate generation.
• Utilized LLMs with few-shot prompting for entity disambiguation.

Main Results:

• Achieved 56% linking accuracy on the MedMentions dataset, outperforming string matching.
• Candidate generation reached 70% recall@5; disambiguation achieved 80% accuracy.
• LLM-generated descriptions did not consistently improve accuracy.

Conclusions:

• LLMs show potential for medical entity linking in low-resource scenarios.
• The proposed method is a lightweight, fine-tuning-free alternative to supervised models.
• The approach is adaptable to other domains and ontologies beyond biomedicine.