UMLS-based data augmentation for natural language processing of clinical research literature

Area of Science:

• Biomedical Natural Language Processing
• Machine Learning
• Data Science

Background:

• Deep learning models in biomedical NLP often suffer from limited training data.
• Existing data augmentation techniques may not fully leverage domain-specific knowledge.
• Improving model performance in low-resource biomedical domains is crucial for advancing research and clinical applications.

Purpose of the Study:

• To develop and evaluate a novel knowledge-based data augmentation method for biomedical NLP.
• To enhance the performance of deep learning models by addressing training data scarcity.
• To investigate the effectiveness of incorporating the Unified Medical Language System (UMLS) knowledge into data augmentation.

Main Methods:

• Extended the easy data augmentation (EDA) method by integrating UMLS knowledge, creating UMLS-EDA.
• Designed experiments to evaluate UMLS-EDA on deep learning architectures for Named Entity Recognition (NER) and classification tasks.
• Compared the performance of UMLS-EDA against baseline models and BERT-based approaches.

Main Results:

• UMLS-EDA significantly improved NER performance for LSTM-CRF models (micro-F1 scores: +5%, +17%, +15%).
• The LSTM-CRF model with UMLS-EDA outperformed BERT transfer learning for NER (0.66 vs. 0.63 micro-F1).
• UMLS-EDA enhanced sentence classification models, achieving a 9% micro-F1 score increase (0.75 to 0.84), surpassing BERT pretraining (0.82).

Conclusions:

• UMLS-EDA is an effective knowledge-based data augmentation method for biomedical NLP.
• The method substantially improves deep learning models for both NER and sentence classification.
• UMLS-EDA offers valuable insights for developing superior deep learning approaches in low-resource biomedical domains.