A workflow utilizing general-purpose large language models for efficient structuring and data mining of bone...

Area of Science:

• Medical Imaging Informatics
• Artificial Intelligence in Oncology
• Clinical Data Management

Background:

• Whole-body bone scintigraphy is crucial for cancer monitoring but reports are unstructured.
• Inefficient data extraction hinders multicenter studies and clinical research.
• Need for automated methods to structure scintigraphy findings.

Purpose of the Study:

• To validate a clinical-logic-guided prompting framework for large language models (LLMs).
• To assess LLM performance in structured information extraction from bone scintigraphy narratives without fine-tuning.
• To evaluate the workflow's efficacy in real-world clinical settings.

Main Methods:

• Established a multicenter ground-truth dataset for benchmarking LLMs.
• Evaluated four LLMs (DeepSeek-R1, DeepSeek-V3, GPT-o3, Gemini 2.5 Pro).
• Implemented a human-in-the-loop workflow with the best-performing LLM (DeepSeek-R1).
• Automated processing of 18,331 patient reports to construct a bone metastasis atlas.

Main Results:

• DeepSeek-R1 showed highest accuracy and stability in structured extraction.
• LLM assistance reduced manual processing time by 74.5%-82.6% with enhanced accuracy.
• Successfully created a bone metastasis atlas for eight common malignancies.

Conclusions:

• Clinical-logic-guided prompt engineering effectively directs general-purpose LLMs for information extraction from bone scintigraphy.
• This low-code paradigm transforms medical narratives into structured, analyzable data.
• Enables efficient data management and empowers large-scale clinical research in oncology.