A Real-Time Clinical Text Information Extractor via LLM.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Oncology

Background:

• Structured data extraction from unstructured clinical text is crucial for oncology research and real-time decision support.
• Automating this process can significantly improve healthcare workflows.

Purpose of the Study:

• To develop and evaluate a modular pipeline using Large Language Models (LLMs) for automated information extraction from Italian gynecologic oncology reports.
• To assess the performance and latency of different LLM architectures for this task.

Main Methods:

• A modular pipeline integrating hierarchical document segmentation, LLM-driven few-shot information extraction, and post-processing was developed.
• Validation employed expert-annotated gold standards and an LLM-as-a-Judge framework.
• Multiple LLM architectures (Gemma3:12b, Gemma3:27b, GPT-oss:20b, Mistral:7b) were evaluated.

Main Results:

• A trade-off between extraction accuracy and computational latency was observed across different LLMs.
• Gemma3:12b demonstrated lower latency and robust performance, suitable for real-time applications.
• GPT-oss:20b showed higher latency, potentially limiting its real-time use.

Conclusions:

• The proposed framework enables rapid and standardized information extraction from clinical reports.
• This offers a scalable solution for integrating structured insights into oncology healthcare workflows.
• LLM selection impacts the balance between performance and real-time applicability.