Agentic Authoring of OMOP Concept Sets from Natural Language

Area of Science:

• Medical Informatics
• Computational Biology
• Observational Research

Background:

• Creating OMOP concept sets from free text is a bottleneck for computable phenotyping.
• Current tools are mainly for expert users, not autonomous agents.

Purpose of the Study:

• To develop an agentic framework for automatically generating OMOP concept sets from free-text descriptions.
• To improve the scalability and efficiency of computable phenotyping in observational research.

Main Methods:

• Combined vocabulary tools, ontology extensions (RxClass, LOINC, Disease Ontology), and procedural guidance.
• Developed an agentic framework for autonomous concept set generation.
• Validated performance using ablation studies, cohort-level EHR data, and a silver-standard benchmark.

Main Results:

• Achieved high performance metrics (Recall@100=0.965, AP@100=0.875) in development set ablation studies.
• Demonstrated excellent cohort-level validation (Precision=0.970, Recall=0.998, Jaccard=0.968).
• Outperformed existing tools and SQL access on an independent benchmark (Recall@100=0.835, AP@100=0.786).

Conclusions:

• The agentic framework effectively automates OMOP concept set generation.
• This approach significantly advances scalable computable phenotyping for observational research.
• Progressive guidance within the framework yielded the best performance.