Alzheimer’s treatment requires early detection; yet, predicting progression is challenging due to significant missing information in medical data for biomarkers and neuroimages. Recent studies tackled the missing data issue in biomarker, by introducing an imputation module to handle the missing values. However, for neuroimaging modalities such as Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET), we still need a reliable system to handle this major issue. To overcome this, we propose an end-to-end hybrid model that is capable of handling both missing biomarker data as well as neuroimages. The proposed model employs a two-fold approach: first, it uses an attention-based multimodal variational autoencoder to impute missing neuroimages and a mask imputation strategy for biomarker data. Second, it leverages a recurrent neural network (RNN) to predict AD progression in future years, effectively handling missing modalities by reconstructing the missing data before making predictions. We performed our experiments on 1369 patients from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset and our model achieved 0.6059 ± 0.0151, 0.6074 ± 0.0163, 0.6166 ± 0.0203, 0.7749 ± 0.0135 in terms of accuracy, precision, recall, and mAUC, respectively. The results confirm that our proposed model can be useful for handling missing data and by utilizing both biomarker and neuroimaging data simultaneously, we can precisely predict the progression of Alzheimer’s disease in clinical settings.