Semi-automated De-identification of German Content Sensitive Reports for Big Data Analytics

Area of Science:

• Medical Informatics
• Data Security
• Natural Language Processing

Background:

• Collaborative research across institutions necessitates robust data security measures.
• Selective de-identification of sensitive information in medical reports is crucial for data sharing.
• The increasing volume of 'Big Data' in healthcare amplifies the need for automated de-identification solutions.

Purpose of the Study:

• To develop and evaluate a semi-automated tool for de-identifying sensitive content in various medical reports.
• To assess the tool's performance in detecting direct and indirect identifiers, medical terms, and filler words.
• To determine the optimal amount of training data required for reliable de-identification.

Main Methods:

• A semi-automated de-identification tool was developed and tested on diverse medical reports (pathology, medical, operation, radiology).
• The tool was evaluated natively and after iterative retraining with increasing sets of manually edited reports (25, 50, 100, 250, 500, 1000).
• Performance was measured by sensitivity and specificity in detecting different categories of sensitive data.

Main Results:

• Native detection rates were 61.3% for direct and 80.8% for indirect identifiers.
• After training with approximately 100 reports, detection rates for direct and indirect identifiers reached 99.5% and 97.2%, respectively.
• Falsely flagged medical terms decreased from 5.3% (native) to approximately 3-4% after training, with minimal flagging of filler words.

Conclusions:

• Continuous training significantly enhances the performance of the developed de-identification tool.
• Training with around 100 edited reports enables reliable detection and labeling of sensitive data across different medical report types.
• While the tool is highly effective, a final review by authorized personnel remains essential for complete data security.