Disambiguating Clinical Abbreviations by One-to-All Classification: Algorithm Development and Validation Study

Area of Science:

• Natural Language Processing
• Biomedical Informatics
• Computational Linguistics

Background:

• Electronic medical records (EMRs) contain valuable patient data, including clinical notes.
• Ambiguous and unstandardized abbreviations in clinical text hinder natural language processing (NLP) for clinical decision support.
• Efficient abbreviation disambiguation is crucial for accurate information extraction from medical records.

Purpose of the Study:

• To enhance the one-to-all (OTA) framework for clinical abbreviation expansion.
• To optimize Bidirectional Encoder Representations From Transformers (BERT) models by developing context-candidate pairs and improving word embeddings.
• To evaluate the efficacy of the enhanced OTA framework in expanding clinical abbreviations using real-world medical data.

Main Methods:

• Utilized three datasets: Medical Subject Headings Word Sense Disambiguation, University of Minnesota, and Chia-Yi Christian Hospital.
• Preprocessed and formatted texts containing polysemous abbreviations for BERT.
• Fine-tuned pretrained models, ClinicalBERT and BlueBERT, generating training and testing dataset pairs.

Main Results:

• BlueBERT achieved high accuracies (e.g., 95.41% macroaccuracy on MSH WSD dataset).
• Demonstrated significant macroaccuracy improvements over baseline models (LSTM, deepBioWSD, Word2Vec+SVM, BioWordVec+SVM).
• Achieved 98.40% macroaccuracy on the University of Minnesota dataset.

Conclusions:

• The enhanced one-to-all (OTA) method effectively disambiguates clinical abbreviations.
• This approach shows potential for improving clinical staff efficiency and research effectiveness.
• Validated the utility of optimized BERT embeddings for clinical abbreviation expansion.