Comparing artificial neural networks to other statistical methods for medical outcome prediction

Survival prediction is important in cancer because it determines therapy, matches patients for clinical trials, and provides patient information. Is a backpropagation neural network more accurate at predicting survival in breast cancer than the current staging system? For over thirty years cancer outcome prediction has been based on the pTNM staging system. There are two problems with this system: (1) it is not very accurate, and (2) its accuracy can not be improved because predictive variables can not be added to the model without increasing the model's complexity to the point where it is no longer useful to the clinician. Using the area under the curve (AUC) of the receiver operating characteristic, the authors compare the accuracy of the following predictive models: pTNM stage, principal components analysis, classification and regression trees, logistic regression, cascade correlation neural network, conjugate gradient descent neural network, backpropagation neural network, and probabilistic neural network. Using just the TNM variables both the backpropagation neural network, AUC.768, and the probabilistic neural network, AUC.759, are significantly more accurate than the pTNM stage system, AUC.720 (all SEs<.01, p<.01 for both models compared to the pTNM model). Adding variables further increases the prediction accuracy of the backpropagation neural network, AUC.779, and the probabilistic neural network, AUC.777. Adding the new prognostic factors p53 and HER-2/neu increases the backpropagation neural network's accuracy to an AUC of .850. The neural networks perform equally well when applied to another breast cancer data set and to a colorectal cancer data set. Neural networks are able to significantly improve breast cancer outcome prediction accuracy when compared to the TNM stage system. They can combine prognostic factors to further improve accuracy. Neural networks are robust across data bases and cancer sites. Neural networks can perform as well as the best traditional prediction methods, and they can capture the power of nonmonotonic predictors and discover complex genetic interactions.<>