De-identification of clinical notes with pseudo-labeling using regular expression rules and pre-trained BERT.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Data Privacy

Background:

• De-identification of clinical notes is crucial for medical research using unstructured text data.
• Limited research exists on de-identifying Korean clinical notes.
• Protecting patient privacy in electronic health records is a growing concern.

Purpose of the Study:

• To develop and evaluate an effective de-identification method for Korean clinical notes.
• To improve the utilization of unstructured clinical text data in Korean medical research.
• To address the gap in de-identification techniques for Korean language medical records.

Main Methods:

• Utilized a large dataset from Seoul National University Bundang Hospital, including radiology and non-radiology notes.
• Employed a two-stage de-identification strategy: rule-based (regular expressions) and semi-supervised machine learning (Korean BERT).
• A rule-based approach was refined using expert-annotated data, then used for pseudo-labeling to fine-tune a pre-trained Korean BERT model.

Main Results:

• The rule-based approach achieved 97.2% precision, 93.7% recall, and 96.2% F1 score on radiology notes.
• The semi-supervised KoBERT-NER model, fine-tuned with pseudo-labeled data, reached 96.5% precision, 97.6% recall, and 97.1% F1 score.
• Validation demonstrated high performance in token-level de-identification for both radiology and non-radiology notes.

Conclusions:

• Combining rule-based and semi-supervised machine learning enhances clinical note de-identification performance.
• The proposed method offers an effective solution for de-identifying Korean clinical text.
• Improved de-identification facilitates secure data sharing and analysis in Korean medical research.