This master thesis work outlines the approach chosen to manage the extraction of information from unstructured and non standardized Italian electronic health records of breast cancer summaries. Our problem translates into a Named Entity Recognition activity that faces completely original classes that do not overlap at all with those normally studied in the literature. Taking into account feasibility and speed, we propose a solution that applies natural language processing procedures to fine-tune a statistical model trained onto a narrow set of annotated data. In order to obtain a gold standard we collaborated with domain experts to manually annotate the datacollection.