Radial Basis Function Neural Networks for the Analysis of Survival Data

The growing interest in artificial neural networks for outcome prediction of oncological patients is motivated by the increasing number of variables related to patient and/or disease characteristics to be investigated and by the possible presence of complex prognostic effects. Neural networks suitable for survival data should consider censoring in a correct way to avoid suboptimal models. Starting from the relationship between survival regression models and generalized linear models with Poisson error, we proposed their extensions as feed-forward neural networks. In particular, radial basis function networks are considered, which can be implemented with standard statistical software. The proposed models can be applied in an exploratory framework, for a single event and in the presence of competing risks. An application of the proposed models to literature data is presented.