Use of neural networks for prediction of graft failure following liver transplantation

Clinical information was gathered prospectively for 295 patients who underwent liver transplantation at the University of Pittsburgh Medical Center, and was divided into sets. The feedforward, fully connected, neural networks had 7 or 8 inputs, a single hidden layer consisting of 3 nodes and a single output node. Using a standard back propagation algorithm, the networks were trained with data from a randomly selected subset of 240 patients while the remaining 55 patients made up the test set. Training was assessed by testing the ability of the network to correctly predict the outcome of the 55 patients in the test set. The accuracy of prediction by the neural network improved each day and so by day 5, 98% of the graft survivors in the test set were correctly predicted while 88% of graft failures in the test set were correctly predicted.