Relation Extraction from Clinical Narratives Using Pre-trained Language Models.

Area of Science:

• Computational linguistics
• Medical informatics
• Artificial intelligence

Background:

• Natural language processing (NLP) is crucial for extracting information from clinical text.
• Traditional methods rely on word embeddings from language models (LMs).
• Pre-trained LMs like BERT excel in open-domain NLP but haven't been applied to clinical relation extraction (RE).

Purpose of the Study:

• To adapt and evaluate Bidirectional Encoder Representations from Transformers (BERT) for clinical relation extraction.
• To assess the performance of fine-tuned BERT models on clinical RE tasks.

Main Methods:

• Developed two distinct implementations of the BERT model tailored for clinical RE.
• Fine-tuned pre-trained BERT language models on clinical datasets for the RE task.

Main Results:

• The fine-tuned BERT models surpassed previous state-of-the-art RE systems.
• Achieved superior performance in two clinical RE shared tasks.
• Demonstrated the effectiveness of LM-based approaches for clinical RE.

Conclusions:

• Pre-trained language models, specifically BERT, show significant promise for clinical relation extraction.
• Fine-tuning BERT models offers a powerful strategy to advance clinical NLP.
• LM-based methods represent a new state-of-the-art for clinical RE tasks.