Sequence Labeling for Disambiguating Medical Abbreviations.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Medical text frequently uses abbreviations to enhance efficiency and protect patient privacy.
• The ambiguity of abbreviations (multiple meanings) poses a significant challenge in clinical documentation.
• A lack of standardized mapping systems complicates the disambiguation of medical abbreviations.

Purpose of the Study:

• To evaluate the effectiveness of sequence labeling techniques for medical abbreviation disambiguation.
• To assess the capability of sequence labeling to handle multiple unique abbreviations within a single text.
• To compare transformer models pre-trained on scientific and medical corpora for this task.

Main Methods:

• Utilized two public datasets for evaluating model performance.
• Compared several transformer models, including those pre-trained on specialized corpora.
• Implemented and assessed sequence labeling approaches against text classification models.

Main Results:

• The proposed sequence labeling approach surpassed the performance of standard text classification models.
• The SciBERT model exhibited robust performance in both sequence labeling and text classification.
• Text classification models achieved comparable performance to sequence labeling only after applying postprocessing filters.

Conclusions:

• Sequence labeling is a feasible and effective method for medical abbreviation disambiguation.
• Transformer models, particularly SciBERT, show significant promise for improving the accuracy of automated clinical note analysis.
• Postprocessing is crucial for enhancing the performance of text classification models in abbreviation disambiguation.