DeIDNER Model: A Neural Network Named Entity Recognition Model for Use in the De-identification of Clinical Notes

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning
• Biomedical Informatics

Background:

• Clinical Named Entity Recognition (NER) is crucial for NLP tasks like information extraction and de-identification.
• Deep learning (DL) methods using word embeddings are increasingly used in clinical NLP.
• Limited research exists on evaluating and combining word embeddings from different domains.

Purpose of the Study:

• To improve NER performance in clinical discharge summaries.
• To develop a DL model that effectively combines various word embeddings.
• To investigate the synergy of standard and contextual embeddings from general and clinical domains.

Main Methods:

• Developed a high-quality, human-annotated corpus of clinical discharge summaries.
• Constructed a NER model incorporating a multi-embedding input layer: standard, contextual, character-level (CNN), and external knowledge embeddings.
• Employed Bidirectional Long Short-Term Memory (Bi-LSTM) and Conditional Random Field (CRF) layers post-embedding.

Main Results:

• The proposed model achieved state-of-the-art performance on public datasets.
• An F1 score of 94.31% was obtained on the internal corpus.
• Incorporating mixed-domain clinically pre-trained contextual embeddings further boosted the internal corpus F1 score to 95.36%.

Conclusions:

• Demonstrated an effective method for combining diverse word embeddings to enhance NER.
• The improved NER performance directly supports and aids downstream clinical note de-identification.
• This approach offers a pathway to more accurate and efficient processing of clinical text data.