Large Language Model Empowered Privacy-Protected Framework for PHI Annotation in Clinical Notes.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Data Privacy

Background:

• De-identifying private health information is essential for confidentiality.
• Existing methods like rule-based and learning-based approaches lack generalizability and require extensive annotated data.
• Large Language Models (LLMs) offer advanced language comprehension but pose privacy risks and high computational demands.

Purpose of the Study:

• To propose LPPA (LLM-powered Privacy-protected PHI Annotation), a novel framework for efficient and private health information annotation.
• To leverage few-shot learning with pre-trained LLMs to generate synthetic clinical notes, minimizing the need for large annotated datasets.
• To enhance privacy protection and PHI annotation accuracy through local fine-tuning of LLMs.

Main Methods:

• Utilizing pre-trained LLMs for few-shot learning to generate synthetic clinical notes.
• Implementing local fine-tuning of LLMs on the generated synthetic data.
• Developing the LPPA framework for privacy-protected PHI annotation.

Main Results:

• Demonstrated effectiveness in PHI annotation accuracy.
• Achieved high efficiency and scalability in the annotation process.
• Successfully generated synthetic clinical notes to reduce data requirements.

Conclusions:

• LPPA offers a robust solution for privacy-protected PHI annotation.
• The framework effectively balances privacy, accuracy, and efficiency.
• LPPA shows significant promise for de-identifying medical records.