Not Fully Synthetic: LLM-based Hybrid Approaches Towards Privacy-Preserving Clinical Note Sharing.

Area of Science:

• Health Informatics
• Medical Data Privacy
• Artificial Intelligence in Healthcare

Background:

• Sharing clinical notes is vital for healthcare research but faces privacy challenges from regulations like HIPAA and GDPR.
• Existing de-identification methods and synthetic data generation often compromise data utility or privacy.
• Limitations include incomplete privacy protection by de-identification and lack of nuance in synthetic notes.

Purpose of the Study:

• To develop a hybrid approach for generating clinical notes that balances patient privacy with data utility.
• To evaluate the effectiveness of this hybrid method in retaining original content and ensuring de-identification.
• To provide a solution for safe and effective sharing of clinical data for research and innovation.

Main Methods:

• A hybrid approach combining de-identification, data filtration, and Large Language Model (LLM) based synthetic note generation.
• Retaining 36%-61% of original clinical note content and using LLM to generate remaining data.
• Evaluating de-identification performance of hybrid notes against standalone methods.

Main Results:

• Hybrid notes retain significant portions of original clinical data (36%-61%) while using LLMs to fill gaps.
• The de-identification performance of hybrid notes meets or exceeds standalone de-identification techniques.
• The proposed method successfully preserves patient privacy and the richness of clinical data.

Conclusions:

• Hybrid clinical notes offer a promising solution for secure and effective clinical data sharing.
• This approach enhances data utility for research without compromising patient privacy.
• The method encourages further innovation in medical data sharing and analysis.