Fine-Grained Prompting in Large Language Models for Accurate and Efficient TNM Staging from Radiology Reports.

Area of Science:

• Oncology
• Medical Informatics
• Artificial Intelligence

Background:

• Accurate TNM staging is crucial for cancer diagnosis and treatment planning.
• Extracting TNM staging information from unstructured radiology reports presents a significant challenge.
• Current large language model (LLM) performance in TNM staging can be limited by prompt complexity.

Purpose of the Study:

• To introduce Fine-Grained Prompting (FGP), a novel approach to enhance LLM performance in TNM staging.
• To improve the accuracy and efficiency of extracting and classifying TNM staging data from radiology reports.
• To evaluate the clinical utility of FGP in real-world cancer staging workflows.

Main Methods:

• Developed FGP by decomposing TNM staging definitions into manageable subtasks.
• Integrated subtask responses to predict the final TNM stage, optimizing prompt length and task simplicity.
• Developed application software integrating FGP for clinician evaluation.

Main Results:

• FGP demonstrated superior performance over basic prompt engineering methods.
• Achieved an 18.5% improvement in T accuracy for lung cancer TNM staging.
• Clinician time efficiency for lung cancer staging more than doubled using FGP-based software compared to manual methods.

Conclusions:

• FGP offers a promising solution for enhancing LLM-based TNM staging from radiology reports.
• The approach significantly improves accuracy and clinical efficiency in cancer staging.
• FGP has the potential to set a new standard for AI-assisted cancer staging, improving patient outcomes.