A neural-Bayesian approach to survival analysis

摘要

Standard survival analysis can be given a neural interpretation in terms of a multi-layered perceptron (MLP) with exponential transfer functions. More hidden units accommodate more complex relationships. The neural interpretation suggests aBayesian analysis, which allows one to introduce sensible priors and to sample from the posterior. We also propose a method for computing p-values from the obtained ensemble of networks, because, in the end, this is the kind of information medical experts are familiar with. We apply our methods on a database regarding patients with ovarian cancer.