Towards Comprehensive Clinical Abbreviation Disambiguation Using Machine-Labeled Training Data.

Area of Science:

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

Background:

• Abbreviation disambiguation in clinical texts is crucial for accurate information extraction.
• Fully supervised methods require expensive, large-scale annotated datasets, limiting their practicality.
• Existing semi-supervised methods often focus on limited abbreviations or lack real-world validation.

Purpose of the Study:

• To evaluate the effectiveness of semi-supervised learning for abbreviation disambiguation in clinical texts.
• To test semi-supervised classification algorithms on a large, real-world dataset of ambiguous medical abbreviations.
• To demonstrate a practical and extensible solution for abbreviation disambiguation in medical NLP.

Main Methods:

• A semi-supervised approach was employed, automatically generating training data by substituting long forms with abbreviations.
• Several semi-supervised classification algorithms were tested.
• The algorithms were evaluated on a large, hand-annotated medical record containing 74 ambiguous abbreviations.

Main Results:

• Classifiers achieved up to 90% accuracy in disambiguating abbreviations.
• The method demonstrated effectiveness despite notable differences between training and test corpora.
• The performance indicates the viability of semi-supervised learning for this task.

Conclusions:

• Semi-supervised abbreviation disambiguation is a viable and extensible approach for medical NLP.
• This method reduces the need for extensive manual data annotation.
• The findings support the adoption of semi-supervised techniques in clinical text analysis.