A Real-Time Clinical Text Information Extractor via LLM

The extraction of structured information from unstructured clinical text is a critical requirement for real-time decision support and research applications in oncology. In this study, we present a modular pipeline leveraging locally deployed Large Language Model (LLM) for the automated extraction of clinically relevant entities from Italian gynecologic oncology reports. The system integrates hierarchical document segmentation, LLM-driven few-shot information extraction and post-processing. Validation is performed combining expert-annotated gold standards with an LLM-as-a-Judge framework. Evaluation across multiple LLM architectures (Gemma3:12b, Gemma3:27b, GPT-oss:20b, Mistral:7b) demonstrates a trade-off between extraction accuracy and computational latency. Normalized runtime analysis indicates that Gemma3:12b achieves lower latency with robust performance, while GPT-oss:20b exhibits higher latency, limiting real-time applicability. The proposed framework facilitates rapid, standardized information extraction from clinical reports, offering a scalable solution for integrating structured insights into healthcare workflows.