Classifying medical relations in clinical text via convolutional neural networks

Highlights：

• This study proposed a CNN-based method (without any external features) for recognizing medical concept relations in clinical records.

• A multi-pooling operation was introduced into the proposed CNN architecture, which aims to capture the position information of local features relative to the concept pair.

• A novel loss function with a category-level constraint matrix was explored.

• The proposed models achieved improved performance compared to previous single-model methods, and the best model is competitive with the ensemble-based method for classifying relations between medical concepts.