Survival models are used in various fields, such as the development of cancertreatment protocols. Although many statistical and machine learning models havebeen proposed to achieve accurate survival predictions, little attention hasbeen paid to obtain well-calibrated uncertainty estimates associated with eachprediction. The currently popular models are opaque and untrustworthy in thatthey often express high confidence even on those test cases that are notsimilar to the training samples, and even when their predictions are wrong. Wepropose a Bayesian framework for survival models that not only gives moreaccurate survival predictions but also quantifies the survival uncertaintybetter. Our approach is a novel combination of variational inference foruncertainty estimation, neural multi-task logistic regression for estimatingnonlinear and time-varying risk models, and an additional sparsity-inducingprior to work with high dimensional data.