Figure 7

Training the W2V-ICD model. Word w₄ occurs in a note that is part of a care episode labeled with the ICD-10 code J21.1. The input node corresponding to w₄ is activated and the error between the output layer prediction and the expected prediction for J21.1 is back-propagated through the network. Same procedure is repeated for J21, with the update scaled by 0.5, and J2 scaled by 0.25, and finally J, scaled by 0.125. When the training is completed, the context vector $\overrightarrow{\mathsf{\text{C}}{w}_4}$ is formed by the weights connecting the input node corresponding to w₄ and the hidden layer of the network.