This research demonstrates how large language models can generate synthetic clinical data with privacy annotations, addressing the critical shortage of training datasets in sensitive domains.

• Domain-adapted LLMs create realistic synthetic clinical texts
• Advanced NER models automatically add de-identification tags
• The synthetic corpora effectively train models for identifying personally identifiable information
• Overcomes privacy barriers that limit healthcare AI progress

For healthcare organizations, this approach provides a path to develop robust de-identification systems without exposing real patient data, accelerating AI adoption while maintaining privacy compliance.

Data-Constrained Synthesis of Training Data for De-Identification