CDR-Agent: Intelligent Selection and Execution of Clinical Decision Rules Using Large Language Model Agents

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Emergency Medicine Decision Support

Background:

• Clinical decision-making in emergency departments (EDs) is complex and time-sensitive.
• Clinical Decision Rules (CDRs) aid diagnosis but are underutilized due to clinician cognitive load.
• Existing tools struggle to integrate CDRs seamlessly into fast-paced clinical workflows.

Purpose of the Study:

• To introduce CDR-Agent, a novel Large Language Model (LLM)-based system for autonomous CDR identification and application.
• To enhance clinical decision-making in emergency settings by reducing the burden of CDR recall and selection.
• To evaluate the performance of CDR-Agent on both synthetic and real-world ED datasets.

Main Methods:

• Development of CDR-Agent, an LLM-powered system to process unstructured clinical notes.
• Creation of two datasets: a synthetic ED dataset and the CDR-Bench dataset for validation.
• Comparative analysis of CDR-Agent against a standalone LLM baseline for CDR selection and overall prediction accuracy.

Main Results:

• CDR-Agent demonstrated significant accuracy gains in CDR selection: 56.3% (synthetic data) and 8.7% (CDR-Bench).
• Overall prediction accuracy improved substantially: 134.0% (synthetic data) and 20.4% (CDR-Bench) compared to the baseline.
• The system also achieved a significant reduction in computational overhead.

Conclusions:

• CDR-Agent effectively enhances the application of Clinical Decision Rules in emergency medicine.
• The LLM-based approach improves diagnostic accuracy and efficiency while reducing computational costs.
• CDR-Agent shows promise for broader application beyond emergency departments in clinical decision support.