Unstructured natural language information contained in electronic health records holds considerable value for health research and innovation. However, manually reviewing these records is impractical and hinders access to the insights they contain. This study introduces a method to automatically extract information from unstructured clinical documents through Domain Entity Recognition and Relation Extraction, then organize the outputs using a domain-specific ontology. The work focuses on oncology, a field that demands significant attention. The primary contribution is a comprehensive methodology that combines multiple resources to achieve this goal. We created a novel annotated dataset of medical evolution notes in Brazilian Portuguese, featuring 1622 documents, 146,769 entities, and 111,716 relations. In the exams domain, the approach reached 78.24% accuracy for entity and relation extraction. Healthcare experts reviewed the method for entity recognition and relation extraction and viewed it favorably, deeming the overall methodology useful for medical practitioners.
