This research leverages LLMs to generate synthetic clinical datasets with pre-annotated personally identifiable information, solving the data scarcity challenge in privacy-sensitive domains.

• Domain-adapted LLMs create realistic clinical texts with embedded PHI markers
• Machine-annotated tags eliminate manual labeling efforts for sensitive information
• Synthetic corpora enable training of de-identification systems without privacy risks
• Practical solution for medical institutions needing quality training data while maintaining compliance

For healthcare organizations, this approach offers a pathway to develop robust de-identification systems without exposing real patient data, potentially accelerating AI adoption in clinical settings while maintaining regulatory compliance.

Data-Constrained Synthesis of Training Data for De-Identification