Using Bayesian networks to predict survival of liver transplant patients

Abstract

The relative scarcity of grafts available for liver transplantation highlights the need to identify patients likely to have good outcomes after treatment. We used transplant information from the United Network for Organ Sharing database to construct a Bayesian network model to predict 90-day graft survival. The final model incorporated a set of 29 pre-transplant variables, and it achieved performance, as measured by area under the receiver operating characteristic curve, of 0.674 by cross-validation and 0.681 on an independent validation set. The results showed a positive predictive value of 91%, while the negative predictive value was lower at 30%. With additional refinement and validation, our model may be useful as an adjunct to clinical experience in identifying patients most likely to have good outcomes following liver transplantation.

Figure 1

The Bayesian network structure created using domain knowledge from clinical expertise and literature review. The nodes are organized with the most important predictors relating to medical status at the top, additional recipient data on the lower left, donor variables on the right, and technical factors in the lower right.

Figure 2

ROC curves for three cross-validation sets and one independent validation set. The AUC for each plot is shown in the lower right, with the 95% confidence interval in parentheses.