A neural-symbolic AI agent system for biomedical concept mapping

Area of Science:

• Biomedical Informatics
• Natural Language Processing

Background:

• Concept mapping links free text to standardized biomedical terms.
• Current methods (rule-based, learning-based) face challenges with ambiguity, scalability, and long-tail concept accuracy.

Purpose of the Study:

• Introduce Medical Concept Mapping (MCM), an agentic workflow using language models.
• Improve accuracy, interpretability, and robustness of biomedical concept normalization, especially for underrepresented and abbreviated concepts.

Main Methods:

• MCM employs language models to rephrase ambiguous mentions into explicit descriptions before concept linking.
• Evaluated across MedMentions, ST21pv, and MCN benchmarks against state-of-the-art baselines.

Main Results:

• MCM achieved superior Recall@1 scores: 63.3 (MedMentions), 60.0 (ST21pv), and 67.9 (MCN).
• Significantly improved zero-shot performance on abbreviated mentions, exceeding baselines by up to 24.8 Recall@1 points.
• Human evaluation showed 79.4% of LLM-generated expansions were reasonable/useful, with GPT-OSS achieving 85.1% approval.

Conclusions:

• MCM offers a more accurate, interpretable, and robust solution for long-tail concept normalization in biomedical applications.
• The agentic workflow effectively addresses limitations of previous concept mapping approaches.