Dealing with censorship in neural network models

Feedforward neural networks have recently been considered as nonlinear tools for modelling survival data. This requires handling censorship, since it is inherent in many such studies, including survival following breast cancer surgery which is the subject of this study. Previous survival studies with neural networks have concentrated on modelling survival beyond a single time interval, or have extended the Cox regression model with separate output nodes for each time interval, often ignoring censorship altogether. There is evidence to suggest that ignoring censored data can introduce significant bias into the estimates of survival. We report a survival model using a traditional MLP architecture, where the issue of censorship has been removed from the network structure and dealt as part of the data structure. We also show two coding methods for missing data which are tested to investigate their effect on the accuracy of survival estimates.