Comparison of proportional hazard model and neural network models in a real data set of intensive care unit patients.

There has been increased interest in using neural network model (NNM) for prognosis tasks. However, the performance of NNM has seldom been compared with that of traditional statistical models such as proportional hazard model (PHM) in real data sets. We conducted a comparative study of PHM and two types of NNM, that is, aggregate single point model (ASPM) and multiple point model (MPM), using a real data set of intensive care unit patients. The three models were developed using the 70% training subset and their predictive accuracy were assessed using the 30% testing subset according to classification accuracy, area under receiver operating curve, and concordance index. Overall, the highest predictive accuracy was found in MPM, followed by PHM and ASPM. MPM is likely to have the potential ability to provide more accurate estimation of prognosis than PHM and ASPM