Electronic health records (EHR) have become an important source of a patient data but characterized by a variety of missing values. Using the variational inference of Bayesian framework, variational autoencoder (VAE), a deep generative model, has been shown to be efficient and accurate to capture the latent structure of complex high-dimensional data. Recently, it has been used for missing data imputation. In this paper, we propose a general framework that incorporates effective missing data imputation using VAE and multivariate time series prediction. We utilize the uncertainty obtained from the generative network of the VAE and employ uncertainty-aware attention in imputing the missing values. We evaluated the performance of our architecture on real-world clinical dataset (MIMIC-III) for in-hospital mortality prediction task. Our results showed higher performance than other competing methods in mortality prediction task.