Clinical validation of an artificial neural network trained to identify acute allograft rejection in liver transplant recipients

Abstract

Artificial neural networks (ANNs) are techniques of nonlinear data modeling that have been studied in a wide variety of medical applications. An ANN was developed to assist in the diagnosis of acute rejection in liver transplant recipients. We investigated the diagnostic accuracy of this ANN on a new data set of patients from the same hospital. In addition, we compared the diagnostic accuracy of the ANN with that of the individual input variables (alanine aminotransferase [ALT] and bilirubin levels and day posttransplantation). Clinical and biochemical data were collected retrospectively for 124 consecutive liver transplantations (117 patients) over the first 3 months after transplantation. Diagnostic accuracy was calculated using receiver operating characteristic (ROC) curve analysis. The ANN differentiated rejection from rejection-free episodes in the new data set over the first 3 months posttransplantation with an area under the ROC curve of 0.902 and sensitivity and specificity of 80.0% and 90.1% at the optimum decision threshold, respectively. The ANN was significantly more specific than ALT or bilirubin level or day posttransplantation at their corresponding optimum decision thresholds (P <.0001). Peak ANN output occurred 1 day earlier than peak values for either ALT or bilirubin (P <.005). The diagnostic accuracy of the ANN was greater than that of any of the individual variables that had been used as inputs. It would be a useful adjunct to conventional liver function tests for monitoring liver transplant recipients in the early postoperative period.