From annotation to adaptation: extracting temporal relations in French clinical narratives

Area of Science:

• Natural Language Processing (NLP)
• Clinical Informatics
• Artificial Intelligence (AI)

Background:

• Automated patient timeline generation requires extracting temporal information from clinical narratives.
• A gap exists in NLP resources for non-English clinical text, specifically French.
• Temporal information extraction is crucial for applications like rare disease diagnosis and care coordination.

Purpose of the Study:

• To present a comprehensive framework for temporal relation extraction from French clinical text.
• To address the lack of NLP resources for French clinical data.
• To evaluate modern AI approaches for this task.

Main Methods:

• Developed specialized annotation guidelines for French medical language.
• Created an annotated corpus of 490 French clinical reports with 12,464 entity-relation pairs.
• Compared transformer-based models, large language models, and parameter-efficient fine-tuning (PEFT), including CamemBERT-bio-base.

Main Results:

• Achieved strong inter-annotator agreement (F1 ≥ 0.94 for core entities).
• PEFT with CamemBERT-bio-base demonstrated the strongest temporal relation extraction performance (F1=0.82-0.87).
• Entity consolidation significantly improved named entity recognition (DATE F1=0.96).

Conclusions:

• Validated temporal relation extraction methods provide a foundation for patient timeline generation systems.
• PEFT offers a computationally efficient and high-performing solution for French clinical NLP.
• Further clinical integration and validation are needed, especially for rare genetic diseases.