The prognosis prediction of cancer patients is important for disease management. We introduce DeepProg, a new computational framework that robustly predicts patient survival subtypes based on multiple types of omic data, using an ensemble of deep-learning and machine-learning models. We apply DeepProg on 32 cancer datasets from TCGA and identified multiple cancer survival subtypes. Patient survival risk-stratification based on DeepProg is significantly better (p-value 7.9e-7 log-rank test) than Similarity Network Fusion based multi-omics data integration in all cancer types. Further comprehensive pan-cancer comparative analysis unveils the genomic signatures common among all the poorest survival subtypes, with genes enriched in extracellular matrix modeling, immune deregulation, and mitosis processes.