MDeID: A dataset for multilingual medical de-identification

Medical Text De-identification (DeID) is a critical process in securing Protected Health Information (PHI) in compliance with regulations such as the Health Insurance Portability and Accountability Act (HIPAA). Multiple benchmark datasets exist for English, but there is a paucity of multilingual datasets. We introduce MDeID, a multilingual dataset formedical text de-identification. It comprises 6034 syntheticclinical notes. MDeID covers 28 PHI categories in seven languages: English, Spanish, Portuguese, French, Italian, Dutch, and German. MDeID encompasses seven clinical note types for each language: discharge summaries, provider notes, diagnosis and health plans, nursing notes, emergency medical services notes, scratch notes, and Subjective, Objective, Assessment and Plan (SOAP) notes. MDeID is synthetically generated, seeded with real-world data, and constructed by enriching rare-category coverage and PHI density. We analyze the behavior of models both trained and evaluated on MDeID and contrast it with experiments using real-world and translated datasets. Our work demonstrates MDeID is an effective benchmark for evaluating multilingual DeID difficulty and dataset-specific transfer patterns across the reported models. MDeID is planned for public release at https://github.com/microsoft/MDeID, with the applicable license and use terms provided in the repository.