Meddies PII Clinical De-identification Model

When clinical AI models train on medical records, the key challenge is removing patient personal information (name, ID, address, etc.) while preserving clinical facts (symptoms, medications, test results, etc.) that are essential for diagnosis.

Meddies PII is an open research model and dataset developed to address this problem.

• Dataset characteristics: Synthetic data generated through dynamic prompting, covering diverse languages, document types, labels, lengths, and text formats.
• Versatility: Aims for stable extraction performance not only on structured forms but also on complex real-world hospital data formats such as nursing notes, JSON/XML exports, and chat-style prompts.

While this model is not a finished, off-the-shelf solution, it provides a useful starting point for balancing data privacy and clinical validity, both of which are necessary for the real-world deployment of medical AI.