OEMA: ontology-enhanced multi-agent collaboration framework for zero-shot clinical named entity recognition

Area of Science:

• Natural Language Processing
• Biomedical Informatics
• Artificial Intelligence

Background:

• Clinical Named Entity Recognition (NER) is vital for extracting structured data from electronic health records (EHRs).
• Traditional supervised NER models require extensive manual annotation, which is costly and time-consuming.
• Existing large language model (LLM)-based zero-shot NER methods face challenges in example selection granularity and prompt-self-improvement integration.

Purpose of the Study:

• To introduce OEMA, a novel zero-shot clinical NER framework.
• To address limitations in example selection and prompt integration in LLM-based zero-shot NER.
• To enhance the accuracy and consistency of clinical NER without manual annotations.

Main Methods:

• OEMA utilizes ontology-enhanced multi-agent collaboration.
• A self-annotator generates candidate examples.
• A discriminator leverages SNOMED CT for clinical relevance filtering.
• A predictor incorporates entity-type descriptions for improved inference.

Main Results:

• OEMA outperforms existing zero-shot baselines on benchmark datasets (I2B2 2010, MTSamples, VAERS) across multiple LLMs.
• Under relaxed-match criteria, OEMA achieves performance comparable to supervised BioClinicalBERT.
• Ontology-based filtering reduces noise and improves semantic alignment, bridging the gap between synthetic and real-world data.

Conclusions:

• OEMA offers a significant advancement in zero-shot clinical NER.
• The framework demonstrates performance approaching supervised methods.
• Future work includes continual learning, open-domain adaptation, and multilingual generalization.