Convolutional neural network for clinical narrative categorization

Stacked or sequential convolutional layers in a Convolutional Neural Network (CNN) have shown state-of-the-art results in Image and Pattern Recognition. Recently, CNN's have shown promising results in Natural Language Processing (NLP) tasks. Using a CNN with concurrent convolutional layers, we conduct text categorization on a clinical narrative dataset with imbalance classes. Clinical narratives are written in natural language, documenting the clinical encounter as observed from the clinician along with the process of care. For this research, we experiment with various CNN architectures with a focus on the embedding layer, the first layer in an NLP-based CNN. The input to the embedding layer is the document matrix and the length is typically determined by a maximum document length. This may not be the best option in the case of highly imbalanced classes. Using simple data analysis, we obtain an optimal document length for the document matrix in the embedding layer of the CNN. Comparing the results from our previous published research on classifying clinical narratives, this CNN architecture provides a significant improvement in the F 1 -Score to our previous ensemble-based approach incorporating multiple methods.