Figure 2:

Knowledge-guided Convolutional Neural Network. Pre-trained word embeddings and CUI embeddings are first downstreamed to a 1-dimensional (1-D) convolutional layer, followed by a max-pooling layer to select the highest value of each word or CUI embedding. Then, those selected values are concatenated with seven structured clinical features. A fully connected hidden layer is further used, followed by a dropout and ReLU activation layer. Finally, another fully-connected layer with softmax function is used to yield the probability of distant recurrence. 5 different combinations of features are implemented