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

With the rapid growth of unstructured clinical narratives in electronic health records (EHRs), clinical named entity recognition (NER) has become a crucial technique for extracting structured medical information. However, traditional supervised models such as CRF and BioClinicalBERT rely on costly manual annotations. Although large language model (LLM)-based zero-shot NER reduces the dependency on labeled data, challenges remain in aligning example selection with task granularity and in effectively integrating prompt design with self-improvement frameworks.

To address these limitations, we propose OEMA, a novel zero-shot clinical NER framework based on ontology-enhanced multi-agent collaboration. OEMA consists of three core components: (1) a self-annotator that autonomously generates candidate examples; (2) a discriminator that leverages SNOMED CT to filter token-level examples based on clinical relevance; and (3) a predictor that incorporates entity-type descriptions to enhance inference consistency and accuracy.

Experimental results on three benchmark datasets, including the real-world I2B2 2010 dataset alongside MTSamples and VAERS, demonstrate that OEMA consistently outperforms existing zero-shot baselines under exact-match evaluation across multiple backbone LLMs (including gpt-3.5, gpt-4.1, and gemini-2.5-flash). Moreover, under relaxed-match criteria, OEMA performs comparably to the supervised BioClinicalBERT model while significantly outperforming the traditional CRF method.

OEMA integrates ontology-guided reasoning with multi-agent collaboration to address two key challenges in zero-shot clinical NER: granularity mismatch and prompt-self-improvement integration. Ablation studies indicate that ontology-based filtering reduces noise and improves semantic alignment, helping to bridge the style gap between synthetic data and real-world clinical narratives.

OEMA advances zero-shot clinical NER and achieves performance approaching supervised models under relaxed-match criteria. Future work will focus on continual learning, open-domain adaptation, negation and assertion status detection, and multilingual generalization to further expand its applicability in clinical NLP.