(A) A graphical representation of the HGM for p: patient, i: lab test, and d: diagnosis data. (B) All graph nodes in (A) have a corresponding vector like those shown in (B). The vector representations can be projected into a shared space with the TransE method, and this projection is optimized for retaining relations in the original data in the embedding via skip-gram optimization. Finally, these vectors are concatenated into the CNN model for mortality prediction.