DataSifterText: Partially Synthetic Text Generation for Sensitive Clinical Notes

Area of Science:

• Health Informatics
• Data Privacy
• Natural Language Processing

Background:

• Electronic health records (EHR) generate vast amounts of unstructured clinical text data annually.
• Secure sharing of this sensitive clinical text is a significant barrier to its full utilization.
• Existing methods for data anonymization often compromise data utility.

Purpose of the Study:

• To introduce DataSifterText, a novel method for generating partially synthetic clinical free-text.
• To enable secure sharing of clinical text data among diverse stakeholders.
• To improve re-identification risk mitigation while maximizing data utility preservation.

Main Methods:

• DataSifterText generates partially synthetic free-text data by disguising the location of true and obfuscated words.
• The method preserves the joint population distribution of the original data.
• Statistical obfuscation techniques are employed to alter word choices, orders, and frequencies.

Main Results:

• Partially synthetic text generated by DataSifterText showed differences comparable to fully synthetic text from previous studies.
• In a CDC work injury case study, 60.9-86.5% of synthetic descriptions clustered with original data, outperforming naive suppression (45.8-85.7%).
• In a MIMIC III case study, synthetic data retained over 80% of original patient health information.

Conclusions:

• DataSifterText offers a robust solution for secure clinical text data sharing.
• The technique effectively balances privacy protection (low identification risk) with high utility preservation.
• Partially synthetic clinical text generated by DataSifterText can be safely utilized by researchers and healthcare providers.