Fig. 1

Framework of PAVE. Given a patient, the event embedding module takes his/her demographics (i.e., age and gender) and medical events plus occurring time \((e_1, t_1), (e_2, t_2) ,\ldots , (e_n, t_n)\) as inputs and generates a sequence of embedding vectors \(q=\{q_1, q_2,\ldots , q_n\} \in R^{n \times k}\). Then three fully connected layers are followed to map q to queries \(Q \in R^{n \times k}\), keys \(K \in R^{n \times k}\) and values \(V \in R^{n \times k}\). Next, a self-attention module is adopted to attend to meaningful patterns between medical events and output attention results \(P= \{P_1, P_2,\ldots , P_n\} \in R^{n \times k}\), which are sent to a pattern attention module to generate the attention result \(h \in R^k\). Finally, a fully connected (FC) layer and Sigmoid layer are leveraged to output the clinical outcome risk