From Report to Record: Prompt-Based Information Extraction from Gynecology Oncology Reports Using LLMs

The growing capabilities of Large Language Models (LLMs) in understanding and generating clinical text are transforming the processing of unstructured medical data. This study presents a prompt-based framework for extracting structured information from Multidisciplinary Tumor Board (MTB) reports on ovarian cancer, written in Italian. Focusing on the anamnestic section, the main features (about Demographic, Scores, Comorbidities, Treatments and Gynecological History) are extracted through in-context learning prompts, and evaluated with the LLM-as-a-judge majority-voting scheme. Results show high accuracy across macroareas and strong agreement among judge models, confirming the robustness of the approach. The occurrence-based analysis further highlights how variable frequency influences extraction reliability, ensuring a more balanced interpretation of model performance. Overall, the framework enables reliable LLM-based information extraction and supports the creation of structured clinical knowledge for oncology applications.