Clinical text classification with rule-based features and knowledge-guided convolutional neural networks.

Area of Science:

• Medical Natural Language Processing
• Artificial Intelligence in Healthcare
• Computational Linguistics

Background:

• Clinical text classification is a core challenge in medical NLP.
• Traditional methods rely on rule-based or knowledge-based feature engineering.
• Deep learning's representation learning capabilities are underexplored in this domain.

Purpose of the Study:

• To develop an effective disease classification method by integrating rule-based features with deep learning.
• To leverage domain knowledge within deep learning models for improved accuracy.
• To address the challenge of classifying clinical text with limited examples.

Main Methods:

• A novel approach combining rule-based features and knowledge-guided deep learning models.
• Identification of critical trigger phrases for disease prediction.
• Training a Convolutional Neural Network (CNN) using word and Unified Medical Language System (UMLS) entity embeddings.
• Utilizing trigger phrases for few-shot learning of disease classes.

Main Results:

• The proposed method was evaluated on the 2008 i2b2 obesity challenge dataset.
• The approach demonstrated superior performance compared to current state-of-the-art methods.
• Achieved high accuracy in clinical text classification tasks.

Conclusions:

• Convolutional Neural Network (CNN) models excel at learning effective hidden features from clinical text.
• Unified Medical Language System (UMLS) Concept Unique Identifiers (CUIs) embeddings significantly enhance clinical text representations.
• Integrating domain knowledge, such as UMLS entities, into CNN models shows significant promise for advancing clinical NLP.