Aplicación de las redes neuronales artificiales a la predicción del resultado a corto, medio y largo plazo del transplante renal pediatrico.

Abstract

Objectives: 1) to develop a model neuronal to identify the evolution in short, medium and large term of the RPT using pre-transplant factors and to contrast it with the logistical regression (LR). 2) to determine the importance of the pre-operative variables, in the result of the RPT. 3) to carry out a computer application of help to the decision to indicate the TRI. Patients, material and method: retrospective study of 271 RPT, live and cadaver donors (first and retrasplants). Variables: donor¿s age, type of donor, time in dialysis, cold isquemia, host¿s age, title of citotoxic antibodies, etiology of the RPT, transplant number for the host, transfusions, incompatibility. Logistic Model: multivariante analysis with Cox¿s regression for 1, 6, 12, 18, 24, 60 and 120 months, analysis of survival with Kaplan Meier¿s, determination of the ROC curves for each moment and calculation of the area under the curve (AUC), cross points and their values: sensibility (Se), specificity (Sp), positive (PPV) and negative (NPV) predictive value, reason of verisimilitude for positive (RVP) and negative (RVN) test. Neural Model: multilayer perceptron with function of activation of hyperbolic tangent, 12 entrance neurons, 1 layer of hidden neurons, 1 of exit; determination of the ROC curves with the same parameters that LR. Realization of the process for the same moments of the LR. Comparison of ROC curves with Hanley and McNeil¿s z. Importance of the variables in the ANN by sensiblity. Results: The results (Se, Sp, PPV, NPV, RVP, RVN) of the ANN are always superior to those of the LR and their differences are always statistically significant [P(1month)=0,0009698, P(6m)=0,000187, P(12m)=4,858E-05, P(18m)=4,383E-05, P(24m)=2,614E-06, P(60m)=1,818E-05, P(120m)=1,511E-05]. The importance of the variables according to the ANN shows the influence of donor's type. The computer program has provided a correct prediction. Conclusions: 1)ANN constitute a tool of prediction of the evolution of the RPT to short, half and large term. 2) the AUC is valid to compare both models. 3) ANN constitutes a potent tool to indicate the TRI. 4) the selected variables are valid to design a neural model of utility in the infantile TR. 5) the superiority of the ANN indicates the existence of non-lineal relationships among the variables.