Disambiguating Clinical Abbreviations Using a One-Fits-All Classifier Based on Deep Learning Techniques.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Clinical abbreviations are essential for concise medical documentation but pose interpretation challenges.
• Misinterpreting abbreviations can negatively impact patient care and clinical support systems.
• Lack of standardization in clinical abbreviations hinders understanding and information extraction.

Purpose of the Study:

• To develop and evaluate a one-fits-all classifier for disambiguating clinical abbreviations using deep contextualized representations.
• To leverage pretrained language models, such as BERT, for improved accuracy in clinical abbreviation sense disambiguation.

Main Methods:

• Experiments involved fine-tuning various clinical BERT models (Bioclinical, BlueBERT, MS_BERT) on a clinical abbreviation dataset.
• A one-fits-all classifier approach was employed to enhance the disambiguation of both common and rare clinical abbreviations.
• Performance was evaluated using accuracy metrics on the University of Minnesota dataset.

Main Results:

• One-fits-all classifiers utilizing deep contextualized representations from Bioclinical, BlueBERT, and MS_BERT achieved high accuracy (98.99%, 98.75%, 99.13%).
• The MS_BERT model demonstrated the highest accuracy at 99.13%, surpassing the previous state-of-the-art (98.39%).
• The proposed method effectively handles rare and unseen abbreviations.

Conclusions:

• Fine-tuning deep contextualized representations from pretrained language models is a robust method for clinical abbreviation disambiguation.
• This approach offers an advantage over building separate classifiers for each abbreviation, improving efficiency.
• Transfer learning with deep learning models facilitates the development of practical clinical abbreviation disambiguation systems.