Korean clinical entity recognition from diagnosis text using BERT.

Area of Science:

• Natural Language Processing
• Medical Informatics

Background:

• Clinical entity recognition typically focuses on electronic health records (EHRs).
• New applications like automatic medical diagnosis require processing diverse text data sources.
• This study addresses the need for extracting Korean clinical entities from a novel dataset sourced from an online medical QA site, distinct from EHRs.

Purpose of the Study:

• To extract Korean clinical entities from a non-EHR medical dataset.
• To evaluate the effectiveness of Bidirectional Encoder Representations from Transformers (BERT) for this task.
• To compare BERT's performance against a benchmark model.

Main Methods:

• Utilized a modified BERT labeling strategy to improve Korean postposition separation.
• Employed a pre-trained multilingual BERT model for initializing the entity recognition model.
• Conducted experiments on a newly constructed clinical entity recognition dataset and a standard NER dataset.

Main Results:

• BERT significantly outperformed the character-level bidirectional LSTM-CRF model across all metrics.
• BERT achieved micro-averaged precision, recall, and F1 scores of 0.83, 0.85, and 0.84, respectively.
• BERT demonstrated superior recall, particularly in detecting out-of-vocabulary (OOV) words.

Conclusions:

• Bidirectional Encoder Representations from Transformers (BERT) with WordPiece tokenization is effective for Korean clinical entity recognition.
• BERT shows superiority over state-of-the-art methods on both new and standard NER datasets.
• This research is among the first to tackle clinical entity extraction from non-EHR data sources.