Algorithm 1.

GRAM Optimization

Randomly initialize basic embedding matrix E, attention parameters ua, Wa, ba, RNN parameter θr, softmax parameters W, b.

repeat

Update E with GloVe objective function (see Section 2.4)

until convergence

repeat

X ← random patient from dataset

for visit Vt in X do

for code ci in Vt do

Refer $\mathcal{G}$ to find ci’s ancestors C′

for code cj in C′ do

Calculate attention weight αij using Eq. (2).

end for

Obtain final representation gi; using Eq. (1).

end for

vt ← tanh(∑i:ci∈Vt gi)

Make prediction ${\widehat{\mathbf{y}}}_t$ using Eq. (4)

end for

Calculate prediction loss $\mathcal{L}$ using Eq. (5)

Update parameters according to the gradient of $\mathcal{L}$

until convergence