A technique for using neural network analysis to perform survival analysis of censored data.

The purpose of this study was to demonstrate how a form of neural network analysis could be used to perform survival analysis on censored data, and to compare neural network analysis with the most commonly used technique for this type of analysis, Cox regression. In this study computer simulated data sets were used. The underlying rules connecting prognostic information to the hazard of death were defined to allow the construction of data sets with specific realistic properties that could be used to demonstrate situations in which neural network analysis had particular strengths in comparison with Cox regression modeling. Using these simulated data sets neural network analysis could produce successful predictive models, find interactions between variables, and recognize the importance of variables that contributed to the hazard rate as a complex function of the variables value and in situations where the proportionality of hazards assumption was violated. It was also demonstrated that neural network analysis was not a 'black box', but could lead to useful insights into the roles played by different prognostic variables in determining patient outcome.