Unsupervised Abbreviation Expansion in Clinical Narratives

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Clinical narratives often use abbreviations and acronyms due to time constraints.
• Accurate expansion of these short forms is crucial for text comprehension and semantic analysis.
• Ambiguous and non-lexicalized abbreviations pose a significant challenge in automated text processing.

Purpose of the Study:

• To develop a completely unsupervised and data-driven algorithm for resolving ambiguous abbreviations in clinical texts.
• To improve the accuracy of abbreviation expansion in German cardiology reports.
• To demonstrate the effectiveness of a domain-independent strategy for abbreviation resolution.

Main Methods:

• Utilized a corpus of 30,000 pseudonymised German cardiology reports.
• Employed a lookup strategy based on word bigrams and unigrams.
• Developed an unsupervised, data-driven algorithm for abbreviation expansion.

Main Results:

• Achieved an F1 score of 0.91 on a test set of 200 text excerpts.
• Demonstrated statistically significant improvement over a baseline approach (p < 0.001).
• Showcased the efficacy of a simple, domain-independent strategy with a large corpus.

Conclusions:

• A simple, domain-independent strategy can effectively resolve abbreviations when a large corpus is available.
• The proposed algorithm enhances text comprehension and semantic processing of clinical narratives.
• Future work includes integrating sentence and abbreviation detection and adapting for acronym resolution.