Résumé
Abstract Genomic medicine relies on precise phenotyping and global data sharing, particularly in the context of rare diseases. However, exchanging medical reports across language barriers remains a major challenge. Manual annotation with the Human Phenotype Ontology (HPO) is also prone to inconsistency and incompleteness, risking the loss of clinically relevant information and potential misdiagnoses. We present ClinFly, an open-source pipeline that automatically de-identifies, translates, and summarizes medical reports into HPO terms, in compliance with health-data privacy standards and FAIR principles. In a multicenter, prospective evaluation, we benchmarked ClinFly against physician annotation for de-identification and phenotypic summarization of 50 non-English medical reports. The method achieved a recall of 99% and precision of 77% for protected health information (PHI) de-identification. For HPO summarization, high-confidence predictions reached a recall of 49% and precision of 92%, improving to 78% recall when including low-confidence terms, with an average of 6.6 HPO terms captured per report. ClinFly enables efficient, automated de-identification of PHI and summarization in HPO format to support genomic medicine across language barriers.