Annotating German Clinical Documents for De-Identification.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Data Management

Background:

• De-identification of clinical documents is crucial for patient privacy.
• German clinical documents present unique linguistic challenges for automated de-identification.
• Existing de-identification methods may not be optimal for the German language.

Purpose of the Study:

• To create robust annotation guidelines for de-identifying German clinical documents.
• To establish a high-quality annotated corpus for training and evaluating de-identification models.
• To develop and baseline an automated de-identification system for German clinical text.

Main Methods:

• Development of specific annotation guidelines for German clinical text.
• Assembly of a corpus of 1,106 discharge summaries and transfer letters.
• Annotation of 44,000 protected health information (PHI) items.
• Training a recurrent neural network (RNN) for automated de-identification.
• Evaluation using inter-annotator agreement (instance and token levels) and F1 scores.

Main Results:

• Achieved high inter-annotator agreement (0.96 instance, 0.97 token level).
• Developed a corpus of 1,106 German clinical documents with 44K annotated PHI items.
• The trained RNN achieved F1 scores exceeding 0.9 for most PHI categories.

Conclusions:

• The developed guidelines and corpus are valuable resources for German clinical document de-identification.
• Automated de-identification using RNNs shows high performance on this dataset.
• The study provides a strong baseline for future research in de-identifying German clinical text.