Automatically disambiguating medical acronyms with ontology-aware deep learning

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare

Background:

• Machine learning (ML) requires extensive labeled data for clinical applications.
• Abbreviation disambiguation is crucial for automated clinical note processing but limited by data scarcity.

Purpose of the Study:

• To develop novel data augmentation techniques for ML-based abbreviation disambiguation.
• To improve model generalization and accuracy in processing clinical notes.

Main Methods:

• Utilized biomedical ontologies for related medical concepts.
• Incorporated global context information from medical notes.
• Trained and tested models on diverse datasets (MIMIC III, CASI, i2b2).

Main Results:

• Achieved up to a 17% increase in abbreviation disambiguation accuracy on hand-labeled data.
• Maintained performance on held-out test sets without sacrificing accuracy.
• Demonstrated improved model generalization through data augmentation.

Conclusions:

• Novel data augmentation techniques significantly enhance ML model performance for clinical abbreviation disambiguation.
• The proposed method addresses data scarcity and imbalance issues in training ML models for healthcare.
• This approach facilitates broader deployment of ML in automated clinical note processing.