In this paper we describe the system developed by the Polimi team for the CRF Filling Shared Task 2026, which focuses on extracting structured variables from clinical notes. The task is challenging due to scarce annotations, heterogeneous clinical language, and the sparsity of the 134 items to be extracted. Our approach relies on prompt-based information extraction using locally deployed open-weight Large Language Models (LLMs). We focused on the Italian subset of the dataset. The pipeline performs zero-shot extraction using task-specific prompts augmented with a glossary of abbreviations derived from unlabeled notes. To improve reliability and reduce hallucinations, the extraction schema is decomposed into multiple prompts targeting groups of variables, whose outputs are merged and refined through deterministic post-processing rules to normalize values and recover missing labels. During development we explored verification stages based on LLM-based prediction validation and synthetic example generation, but these strategies did not improve performance and were not included in the final system. On the development set, the best configuration based on Mistral Small 3.2 24B Instruct achieved an F1-score of 67.51%. On the official test set, our system ranked third overall and second among systems evaluated on the Italian subset, achieving an F1-score of 63%.

In recent years, it has become common for patients to get full access to their Electronic Health Records (EHRs), thanks to the advancements in the EHRs systems of many healthcare providers. While this access empowers patients and doctors with comprehensive and real-time health information, it also introduces new challenges, in particular due to the unstructured nature of much of the information within EHRs. To address this, we propose a pipeline to structure clinical notes, providing them with a clear and concise overview of their health data and its longitudinal evolution, also allowing clinicians to focus more on patient care during consultations. In this paper, we present preliminary results on extracting structured information from anamneses of patients diagnosed with ST-Elevation Myocardial Infarction from an Italian hospital. Our pipeline exploits text classification models to extract relevant clinical variables, comparing rule-based, recurrent neural network and BERT-based models. While various approaches utilized ontologies or knowledge graphs for Italian data, our work represents the first attempt to develop this type of pipeline. The results for the extraction of most variables are satisfactory (f1-score > 0.80), with the exception of the most rare values of certain variables, for which we propose future research directions to investigate.