An Efficient and Reliable Agent-Based System for Clinical Data Governance.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare

Background:

• Clinical data governance is crucial for reliable intelligent healthcare systems.
• Real-world clinical records often have complex data quality issues, challenging existing methods.
• Directly using Large Language Models (LLMs) can lead to hallucinations and inefficiency.

Purpose of the Study:

• To propose GovernAgent, an LLM-driven framework to address limitations in clinical data governance.
• To improve semantic fidelity and processing efficiency in resolving clinical data anomalies.
• To enhance the reliability and applicability of intelligent healthcare systems.

Main Methods:

• A hierarchical governance mechanism with cascading Note- and Section-Level agents.
• A Constrained Action Planning mechanism with a hybrid "Copy-Generate" action space.
• Utilizing LLMs to systematically disentangle and resolve multi-level data quality issues.

Main Results:

• GovernAgent demonstrated improved governance accuracy and efficiency on real-world hospital datasets.
• The framework significantly minimized hallucinations and preserved medical provenance.
• High practical adaptability and empowerment of downstream clinical applications were observed.

Conclusions:

• GovernAgent offers an effective LLM-driven solution for complex clinical data governance challenges.
• The hierarchical and constrained approach mitigates LLM limitations like hallucinations and computational bottlenecks.
• This framework enhances the quality and utility of clinical data for intelligent healthcare applications.