Combining data augmentation and domain information with TENER model for Clinical Event Detection

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Biomedical Informatics

Background:

• Deep learning for clinical information extraction is a growing trend.
• Clinical Event Detection (CED) faces challenges due to complex medical terminology and data scarcity.
• Existing deep learning models show limitations in recognizing obscure terms and maintaining robustness.

Purpose of the Study:

• To improve the performance of Clinical Event Detection (CED).
• To address the challenges of obscure medical terms and limited datasets in CED.
• To introduce a novel framework combining data augmentation and domain information.

Main Methods:

• Proposed a multi-granularity information fusion encoder-decoder framework using the TENER model for CED.
• Utilized the BioBERT pre-trained language model for word-level feature generation.
• Developed a new data augmentation method for sequence labeling tasks.

Main Results:

• Achieved an F1-score of 80.26% on the 2012 i2b2 challenge dataset.
• Obtained a type accuracy of 93% and a Span F1-score of 90.33%.
• Outperformed existing state-of-the-art approaches in Clinical Event Detection.

Conclusions:

• The proposed framework effectively addresses challenges in CED, including recognition of professional terms and data scarcity.
• The integration of BioBERT and a novel data augmentation method enhances model performance and robustness.
• The study demonstrates the successful application of the TENER model to CED, setting a new benchmark.