Classifying medical relations in clinical text via convolutional neural networks

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Deep learning models demonstrate strong performance in general domain relation classification.
• Accurate relation classification in clinical records is crucial for medical informatics.
• Existing methods may rely on external features, limiting generalizability.

Purpose of the Study:

• To propose a novel convolutional neural network (CNN) architecture for medical relation classification.
• To explore a new loss function incorporating a category-level constraint matrix.
• To evaluate the model's performance on clinical records without external feature dependencies.

Main Methods:

• Development of a CNN architecture with a multi-pooling operation.
• Implementation of a category-level constraint matrix loss function.
• Experimental evaluation using the 2010 i2b2/VA relation corpus.

Main Results:

• The proposed CNN model with multi-pooling achieved superior performance compared to previous single-model methods.
• The model demonstrated effectiveness without relying on any external features.
• The best performing model showed competitive results against existing ensemble-based methods.

Conclusions:

• The novel CNN architecture and loss function are effective for medical relation classification.
• The approach offers a robust solution for extracting relations from clinical records.
• This method provides a competitive alternative to existing, more complex, or feature-dependent systems.