An Efficient Method for Deidentifying Protected Health Information in Chinese Electronic Health Records: Algorithm...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Data Privacy

Background:

• Electronic health records (EHRs) in China offer potential for medical research but contain sensitive protected health information (PHI).
• Direct use of EHR data raises privacy concerns, necessitating effective deidentification methods.
• Existing deidentification techniques struggle with limited Chinese EHR data and complex language features.

Purpose of the Study:

• To propose a novel deep learning model for Chinese PHI deidentification, overcoming overfitting and data scarcity issues.
• To enhance the utility of digitalized EHR data for real-world medical research while ensuring patient privacy.

Main Methods:

• A hybrid model combining TinyBERT for feature extraction and Conditional Random Field (CRF) for prediction was developed.
• A data augmentation strategy, integrating sentence generation and mention replacement, was employed to address insufficient training data.
• The model was evaluated on collected Chinese EHRs.

Main Results:

• The proposed method achieved superior performance compared to five BERT-based baseline methods.
• Achieved high metrics: microprecision of 98.7%, micro-recall of 99.13%, and micro-F1 score of 98.91%.
• Demonstrated a 40% increase in efficiency over baseline methods.

Conclusions:

• The TinyBERT-based model significantly improves the performance and efficiency of Chinese PHI deidentification.
• The hybrid data augmentation effectively addresses the challenge of limited Chinese EHR data.
• The method facilitates the secure utilization of EHR data for advancing medical research.