Performance of neural nets, CART, and Cox models for censored survival data

This paper illustrates the relative performance of several techniques for nonlinear predictive modeling of simulated censored clinical survival data on the basis of measured risk factors: a neural net approach developed in our group, the CART technique, and the Cox model with (and without) quadratic interactions. Simulated follow-up data is first generated by combining empirical multivariate distributions of clinical factors in breast cancer patients with hypothetical nonlinear risk structures, which are thus `known'. The performance of these analysis methods is evaluated here by comparing the `known' and predicted scores on training and validation (generalization) samples containing 500 patients each. The neural net has the best performance for a complex risk structure in which three-factor interactions play an important role.