Generalizable multilingual medical text anonymization using generative instruction tuning

Area of Science:

• Medical Informatics
• Natural Language Processing
• Data Privacy

Background:

• High-quality medical data is crucial for research but contains sensitive patient information.
• Current anonymization methods are domain-specific, require manual data, and are difficult to scale.
• A scalable, privacy-preserving solution is needed for utilizing unstructured clinical text.

Purpose of the Study:

• To develop a reproducible, annotation-free framework for training and adapting LLM-based medical text anonymization models.
• To enable privacy-preserving use of medical text across diverse settings and languages.
• To reduce reliance on manual annotation and real patient data.

Main Methods:

• Developed a generative medical anonymization model using synthetic data and instruction tuning of generative LLMs.
• Created an annotation-free framework for training and adapting models.
• Evaluated performance on synthetic datasets and real-world patient requests, assessing accuracy, recall, precision, and meaning preservation.

Main Results:

• Generative models trained with the synthetic framework outperformed baseline systems across multiple medical domains.
• Models achieved high accuracy in anonymizing sensitive information and high fidelity in preserving non-sensitive text.
• The framework demonstrated effectiveness with small datasets, generalization to unseen fields, and multilingual support without additional training.

Conclusions:

• The study presents a reproducible, annotation-free approach for effective medical text anonymization.
• This framework reduces the need for real patient data and lowers adaptation costs.
• It facilitates broader use of unstructured clinical information for research and service improvement.