Abstract
Artificial neural networks applied to the survival analysis: A comparative stud y with Cox regression model in its predictive aspect. The purpose of this study was to compare the performance in prediction between the models of Artificial Neural Networks (ANN) and Cox proportional hazards models in the context of survival analysis. More specifically, we tried to verify: a) if the model of hierarchical neural networks is more accurate than Cox's model, and b) if the model of sequential neural networks signifies an improvement with respect to the hierarchical neural networks model. The accuracy was evaluated through resolution (the area under the ROC curve) and calibration (Hosmer-Lemeshow test) measures using survival data. Results showed that hierarchical neural networks outperform Cox's model in resolution while sequential neural networks do not suppose an improvement with respect to hierarchical neural networks. F inally, ANN models produced survival curves that were better adjusted to reality than Cox's model.
