Figure 5.

Training the RI-ICD model. Word w₄ occurs in a note that is part of a care episode labeled with the ICD-10 code J21.1. Its context vector $\overrightarrow{\mathsf{\text{C}}{w}_4}$ is therefore updated by adding the index vector for the code J21.1. This context vector is constructed from the weighted sum of index vectors of its parts: $\sum \overrightarrow{\mathsf{\text{I}}}=\left(0.125\times \stackrel{⃗}{{\mathsf{\text{I}}}_{\mathsf{\text{J}}}}\right)+\left(0.25\times \stackrel{⃗}{{\mathsf{\text{I}}}_{\mathsf{\text{J}}2}}\right)+\left(0.5\times \stackrel{⃗}{{\mathsf{\text{I}}}_{\mathsf{\text{J}}21}}\right)+\left(1.0\times \stackrel{⃗}{{\mathsf{\text{I}}}_{\mathsf{\text{J}}21.1}}\right)$. As a result, w₄'s context vector becomes more strongly associated with the code J21.1 and - to a lesser degree - with all superclasses of J21.1 in the ICD-10 taxonomy. The same update is applied to the context vectors of all other words in care episodes labeled as J21.1.