w_embed = embed_layer(input_w)
v_embed = embed_layer(input_v)
bias_layer = Embedding(input_dim=len(code_cat), output_dim=1, name='bias')
w_bias = bias_layer(input_w)
v_bias = bias_layer(input_v)
inner = Multiply()([w_embed, v_embed])
inner = Lambda(lambda x: K.sum(x, axis=-1, keepdims=True))(inner)
merged = Concatenate(axis=-1)([inner, w_bias, v_bias])
output_cooccur = Lambda(lambda x: K.sum(x, axis=-1, keepdims=False),
name='cooccur')(merged)
input_DX = Input(shape=(len(DXs), ))
DX_embed = embed_layer(input_DX)
DX_embed = MaskedDense(md_width, activation='relu')(DX_embed)
DX_embed = MaskedSum()(DX_embed)
input_PR = Input(shape=(len(PRs), ))
PR_embed = embed_layer(input_PR)
PR_embed = MaskedDense(md_width, activation='relu')(PR_embed)
PR_embed = MaskedSum()(PR_embed)
input_other = Input(shape=(8, ))
merged = Concatenate(axis=1)([DX_embed, PR_embed, input_other])
merged = Dense(fc_width, activation='relu')(merged)
merged = Dropout(dropout)(merged)
output_readm = Dense(2, activation='softmax', name='readm')(merged)
model = Model(inputs=[input_DX, input_PR, input_other, input_w, input_v],
outputs=[output_readm, output_cooccur])
adam = Adam(lr=lr)
model.compile(loss={
model = Model(inputs=[input_DX1, input_DX, input_PR, input_hosp, input_other], outputs=prediction)
if model_name=='att_lr':
input_DX1 = Input(shape=(1,))
DX1_embed = Embedding(input_dim=len(code_cat), output_dim=code_embed_dim, mask_zero=True, embeddings_initializer=embed_initializer,
name='DX1_embed')(input_DX1)
input_DX = Input(shape = (len(DXs),))
DX_embed = Embedding(input_dim=len(code_cat), output_dim=code_embed_dim, mask_zero=True, embeddings_initializer=embed_initializer,
name='DX_embed')(input_DX)
input_PR = Input(shape = (len(PRs),))
PR_embed = Embedding(input_dim=len(code_cat), output_dim=code_embed_dim, mask_zero=True, embeddings_initializer=embed_initializer,
name='PR_embed')(input_PR)
code_embed = Concatenate(axis=1)([DX1_embed, DX_embed, PR_embed])
for j in range(n_att_layers):
code_embed = MultiHeadAttention(head_num=n_heads, use_bias=False, activation='linear')(code_embed)
code_embed = MaskedSum()(code_embed)
input_hosp = Input(shape=(1,))
hosp_embed = Embedding(input_dim=len(hosp_cat), output_dim=hosp_embed_dim, input_length=1)(input_hosp)
hosp_embed = Reshape((hosp_embed_dim, ))(hosp_embed)
input_other = Input(shape=(other_mat_train.shape[1], ))
merged = Concatenate(axis=1)([code_embed, hosp_embed, input_other])
prediction = Dense(2, activation='softmax')(merged)
model = Model(inputs=[input_DX1, input_DX, input_PR, input_hosp, input_other], outputs=prediction)
if model_name=='att_sum_lr':
input_DX1 = Input(shape=(1,))
DX1_embed = Embedding(input_dim=len(code_cat), output_dim=code_embed_dim, mask_zero=True, embeddings_initializer=embed_initializer,
name='DX1_embed')(input_DX1)
input_DX = Input(shape = (len(DXs),))
DX_embed = Embedding(input_dim=len(code_cat), output_dim=code_embed_dim, mask_zero=True, embeddings_initializer=embed_initializer,
name='DX_embed')(input_DX)