Leveraging Large Language Models for Clinical Abbreviation Disambiguation.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Clinical Data Analysis

Background:

• Clinical abbreviation disambiguation is essential for medical information retrieval.
• Existing methods face challenges with limited data and ambiguous abbreviations.
• Accurate interpretation of abbreviations is vital for clinical text analysis.

Purpose of the Study:

• To improve clinical abbreviation disambiguation using Large Language Models (LLMs) and data augmentation.
• To address challenges of limited instances and ambiguous interpretations in clinical texts.
• To leverage BIOGPT for generating contextually relevant data to enhance disambiguation accuracy.

Main Methods:

• Utilized BlueBERT and Transformers for contextual understanding within LLMs.
• Employed a Generative Model (BIOGPT) pretrained on biomedical literature for data augmentation.
• Generated diverse, contextually relevant instances of clinical text for abbreviation expansion.
• Evaluated the approach on the CASI dataset, partitioned into training, validation, and test sets.

Main Results:

• Data augmentation with the Generative Model significantly improved disambiguation performance.
• Enhanced performance was particularly notable for abbreviation senses with limited instances.
• The method effectively addressed dataset imbalance and challenges from similar clinical concepts.
• The proposed model achieved good accuracy on the test set, outperforming previous methods.

Conclusions:

• LLMs and generative techniques are effective for clinical abbreviation disambiguation.
• Data augmentation using BIOGPT successfully addresses data scarcity and ambiguity.
• The approach offers a robust solution for improving the accuracy of clinical text analysis.