Temporal Annotation of German Clinical Language in Real and Synthetic Clinical Documents: Corpus Development and...

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Computational Linguistics

Background:

• Temporal information is crucial for clinical decision-making.
• Automatic extraction of temporal data requires annotated clinical reports.
• German clinical language resources for temporal tagging are scarce.

Purpose of the Study:

• Develop a TimeML-compliant annotation schema for German medical language.
• Adapt existing English guidelines for temporal entities and relations.
• Train baseline temporal taggers for German clinical documents.

Main Methods:

• Adapted English temporal annotation guidelines for German clinical jargon.
• Utilized 5 clinically trained annotators for refinement.
• Annotated two German corpora (3000PA_J and GraSCCo) using TimeML standards.
• Developed Bidirectional Encoder Representations from Transformers (BERT)-based taggers.

Main Results:

• Created 3000PA_J-temp (non-distributable) and GraSCCo-temp (publicly available) corpora.
• Achieved high interannotator agreement (IAA) F₁-scores (0.9) for temporal named entity recognition.
• Temporal relation extraction IAA F₁-scores were 0.57 (GraSCCo) and 0.41 (3000PA_J).
• Baseline tagger performance reached F₁-scores of 0.64-0.85 for named entities and 0.60-0.64 for relation extraction.

Conclusions:

• Introduced the first TimeML-compliant annotation scheme for German clinical language.
• Developed the first publicly accessible, temporally annotated German clinical corpus (GraSCCo-temp).
• Trained the first TimeML-compliant time tagger for German clinical text.
• Generated substantial temporal metadata, comparable to large English clinical datasets.