Conditional random fields for clinical named entity recognition: A comparative study using Korean clinical texts.

Area of Science:

• Natural Language Processing
• Health Informatics
• Clinical Data Mining

Background:

• Electronic Health Record (EHR) systems are increasingly adopted globally.
• Technologies for processing unstructured clinical text within EHRs are still developing.
• Clinical Named Entity Recognition (NER) is crucial for extracting information from these texts.

Purpose of the Study:

• To evaluate conventional NER methods: dictionary-lookup string matching and Conditional Random Fields (CRFs).
• To assess the performance of these methods on clinical texts from South Korean rheumatism patients.
• To identify the most effective approach for clinical NER in this context.

Main Methods:

• Utilized discharge summaries from 200 rheumatic patients at Seoul National University Hospital.
• Applied dictionary-lookup string matching and CRFs to identify semantic types in clinical notes.
• Employed 10-fold cross-validation to evaluate model performance.

Main Results:

• CRFs achieved higher performance than string matching for most semantic types (median F1=0.761).
• String matching was more effective for identifying hospital visit information.
• CRFs demonstrated robust performance (median F1=0.811) even with basic features.

Conclusions:

• CRFs are a suitable method for clinical NER in Korean clinical narrative documents.
• Enhancing training data and feature engineering can further improve accuracy.
• Automated patient history summarization is a potential future application.