De-Identifying Swedish EHR Text Using Public Resources in the General Domain

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Developing de-identification models for electronic health records (EHR) typically requires sensitive patient data, posing privacy risks.
• Existing methods often struggle with low-resource languages and sensitive data constraints.

Purpose of the Study:

• To investigate the effectiveness of incorporating non-sensitive public datasets into EHR training data for de-identification.
• To improve the precision and recall of de-identifying Swedish clinical notes.

Main Methods:

• A deep learning model utilizing recurrent neural networks was trained on Swedish EHR data augmented with scientific medical text and Wikipedia word vectors.
• The model's performance was evaluated on pseudonymized Swedish EHR clinical notes.

Main Results:

• Precision improved from 55.62% to 85.01%, and recall improved from 80.02% to 87.15% after adding Wikipedia word vectors.
• The enhanced model demonstrated superior performance in de-identifying Swedish clinical text.

Conclusions:

• Non-sensitive general domain text, such as Wikipedia word vectors, can effectively train robust de-identification models for clinical text.
• This approach offers a viable solution for de-identifying sensitive data in low-resource languages, enhancing patient privacy.