The balanced application of a model for the estimate of outcomes of liver transplantation, in concert with assessment of disease severity, would not only improve transplant outcomes and maximize patient benefit from transplantation, but also facilitate informed decision making by patients and their relatives when considering transplantation. So far, however, linear discriminating methods have failed to attain sufficient power to predict post-transplant prognosis. Therefore, our aim was to develop a cancer-specific prognostic model by a nonlinear methodology based on pretransplant characteristics. With data collected retrospectively from 290 liver transplant recipients with HCC from February 1999 to August 2009, a multilayer perceptron (MLP) neural network was constructed to predict mortality risk after transplantation. Its predictive performances at posttransplant 1-, 2-, and 5-year intervals were evaluated using a receiver operating characteristic curve. By the forward stepwise selection in MLP network, donor age, donor body mass index, recipient hemoglobin, serum concentrations of total bilirubin, alkaline phosphatase, creatinine, aspartate aminotransferase, international normalized ratio of prothrombin time, and Na(+); alpha fetoprotein categorization, total diameter, number of tumor lesions, presence of imaging macrovascular invasion, and lobe distribution of the tumor were identified to be the optimal input features. The MLP, employing 24 inputs and 7 hidden neurons, yielded c-statistics of 0.909 (P < .001) in the 1-year, 0.888 (P < .001), in the 2-year, and 0.845 (P < .001) in the 5-year prediction. Post-transplant prognosis is a multidimensional, nonlinear problem, and the specific MLP can achieve high accuracy in the prediction of posttransplant mortality risk for HCC recipients. The pattern recognition methodologies like MLP hold promise for solving outcome prediction after liver transplantation.
Read full abstract