End-to-End Clinical Validation of a Human-Supervised Large Language Model Agent for Enterprise Surgical Pathology...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Pathology

Background:

• Large language models (LLMs) show potential for pathology tasks but often lack clinical validation or operate outside secure environments.
• Existing LLM applications in healthcare are frequently experimental and may not meet regulatory compliance standards.

Purpose of the Study:

• To develop and clinically validate a rule-guided, agent-based LLM for assisting gastrointestinal (GI) biopsy reporting.
• To automate report structuring and enhance efficiency while maintaining pathologist control over diagnosis.

Main Methods:

• Developed a rule-guided, agent-based LLM (Microsoft 365 Copilot) within a HIPAA-compliant environment.
• Configured the agent with a fixed rule-based prompt and a quick-text knowledge base.
• Conducted a prospective validation on 94 GI biopsy cases, evaluating accuracy, safety, and workflow integration.

Main Results:

• The AI agent achieved 100% accuracy in identifying organ, sub-organ, and procedure context, and accurate shorthand expansion.
• Minor formatting deviations occurred in 8.5% of cases, and minor diagnostic misinterpretations in 2% without hallucinated diagnoses.
• AI-assisted reporting was significantly faster (33-37 second reduction) and showed lower variability compared to manual methods.

Conclusions:

• A constrained, agent-based LLM can safely enhance pathology reporting efficiency and consistency for GI biopsies.
• The system successfully integrated into the clinical workflow, preserving diagnostic authority with the pathologist.
• LLMs are best suited for clerical augmentation, requiring careful pathologist review due to inherent stochastic errors and variability.