We examined whether we could develop models based on data provided to the United States Renal Data System (USRDS) to accurately predict survival. Records were obtained from patients beginning dialysis in 1990 through 2007. We developed linear and neural network models and optimized the fit of these models to the actual time to death. Next, we examined whether we could accurately predict survival in a dataset containing censored and uncensored patients. The results with these models were contrasted with those obtained with a Cox proportional hazards model fit to the entire dataset. The average C statistic over a 6-month to 10-year time range achieved with these models was approximately 0.7891 (linear model), 0.7804 (transformed dataset linear model), 0.7769 (neural network model), 0.7774 (transformed dataset neural network model), 0.8019 (Cox model), and 0.7970 (transformed dataset Cox model). When we used the Cox proportional hazards model, superior C statistic results were found at time points between 2 and 10 years but at earlier time points, the Cox model was slightly inferior. These results suggest that data provided to the USRDS can allow for predictive models which have a high degree of accuracy years following the initiation of dialysis.

We examined machine learning methods to predict death within six months using data derived from the United States Renal Data System (USRDS). We specifically evaluated a generalized linear model, a support vector machine, a decision tree and a random forest evaluated within the context of K-10 fold validation using the CARET package available within the open source architecture R program. We compared these models with the feed forward neural network strategy that we previously reported on with this data set.