Clinical entity-aware domain adaptation in low resource setting for inflammatory bowel disease

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Clinical Data Mining

Background:

• Electronic health records (EHRs) contain vast amounts of unstructured clinical text.
• Extracting valuable information from this text is challenging and requires automated tools.
• Biomedical entity extraction is crucial for medical research and patient care.

Purpose of the Study:

• To develop a robust biomedical entity extraction tool for inflammatory bowel disease (IBD) data.
• To adapt a general language model for efficient IBD-related information extraction.
• To create a comprehensive pipeline for clinical Dutch named entity recognition (NER).

Main Methods:

• Domain-specific pre-training and entity-aware masking strategies were employed.
• Contrastive learning was used to fine-tune a general language model.
• An end-to-end NER tool was developed with annotation, correction, and visualization features.

Main Results:

• A novel pipeline for clinical Dutch NER was successfully developed.
• The approach demonstrated efficient domain adaptation for IBD information extraction.
• The developed NER tool streamlines medical record curation and clinical workflows.

Conclusions:

• The research presents a significant advancement in automated clinical text analysis.
• The developed tool enhances the retrieval of crucial clinical concepts from unstructured EHR data.
• This work supports improved medical record curation and clinical decision-making for IBD.