Meddies improves clinical data privacy with multilingual PII detection
Meddies-PII: A Multilingual Framework for Personally Identifiable Information Extraction in Clinical De-identification
Computation and LanguageArtificial Intelligence
Summary
Protecting patient privacy in medical records requires identifying personal information correctly. Creating real training data for this is expensive, and previous synthetic alternatives were less detailed. To solve this, the authors made a huge set of one million fake clinical documents in 17 languages, carefully checked for accuracy. They trained a model to find personal details in these documents and found it outperformed other methods across many test sets. They will share their dataset, model, and tools publicly for others to use.
What this means in practice
- •For hospital data teams: Improve automatic removal of personal information from medical records in many languages using a large, validated synthetic dataset and trained model.
- •For healthcare software developers: Develop multilingual data privacy tools for clinical applications using a publicly released model and framework tailored for PII extraction.$Commercial implications: Enables creation of privacy-compliant healthcare software products that automatically detect PII across numerous languages.
Authors
Linh Uyen Le, Christian Hoang, Huy Hoang Ha
Abstract
Clinical de-identification relies on accurately identifying personally identifiable information (PII). However, manually annotated datasets are costly to construct, while existing synthetic alternatives often provide limited details about their generation process or rely on relatively simple synthesis strategies. We introduce Meddies-PII-Dataset, a corpus of one million synthetic clinical documents spanning seventeen languages and nine PII labels. The documents are generated using attribute-conditioned prompts and validated through thirteen deterministic gates that enforce structural and annotation consistency. To evaluate the dataset's utility, we train Meddies-PII-Model, a BIOES token classifier, and compare it with existing PII extraction systems using exact-match entity-level F1. Meddies-PII-Model achieves the highest performance among the evaluated systems on all reported benchmarks, with a mean F1 of 0.827 across fifteen external benchmarks, compared with 0.658 for the strongest baseline. Upon acceptance, we will publicly release the dataset, benchmark suite, model, generation framework, and evaluation code to support research on multilingual clinical de-identification.