A comparison of regression methods with artificial neural networks for the prediction of the outcome of patients on chronic dialysis

grantor: University of TorontoChronic dialysis patients suffer from a high mortality rate. The ability of multivariable regression models and artificial neural networks (ANNs) to predict the survival of individual chronic dialysis patients was assessed. The data consisted of 1381 incident dialysis cases. Logistic and Cox regression models and feed-forward multilayer perceptron ANNs were derived on 1019 cases and then evaluated on the remaining 299 cases. Older age, white race, diabetes, vascular disease, CHF, lower albumin, and peritoneal dialysis modality were significantly associated with decreased survival in univariable and multivariable analyses. The AUROC ± SE for the Cox and ANN models were 0.626 ± 0.032 and 0.877 ± 0.020 respectively ([chi]2 = 50.44, p < 0.0001). The crude classification rates were 62.1 and 80.1% respectively (p = 0.012). In the current dataset, an ANN had superior predictive performance for the survival of chronic dialysis patients compared to a standard regression model. Confirmation of this finding in a new dataset is required.M.Sc