Prediction of outcome of renal transplant survival, application of artificial neural networks

ABSTRACT

One of the purposes of the evolving of the field of medical informatics is to develop decision support systems that enhance the human ability to diagnose, treat, and assess prognoses of pathologic conditions. This work is to compare the ability of an artificial neural network and a model of logistic regression to predict individual survival status at 2 years after renal transplant. Artificial neural networks [ANN s] are new computational tools, which once trained, can recall proper outputs for a specific set of inputs never encountered before. Between 1976 and 1997, 1000 patients with End-Stage-Renal Disease [ESRD] were subject to renal transplant. Survival status at 2 years was known in 725 patients, while censored cases with less than that period were excluded. Logistic regression model was built and a neural network was trained on randomly selected 80% of patients [580 patients] to predict individual status at 2 years [status = "1" for "graft loss" and "O" for "graft survival"]. We classified the risk factors into pretransplant, transplant [technical], and post-transplant predictors. The performance of the LR and ANN models, revealed a sensitivity [percentage of correctly predicted deaths] of 10.6% and 87.6%, a specificity [percentage of correctly predicted survivors] of 99% and 84%, with an overall accuracy of 85.3% and 85.8% respectively. The results show that neural network has a higher accuracy in predicting the sensitivity at the 2-years survival status. It has also a better balance between the correct prediction of losses and survivors. Probably, still some new markers are needed to differentiate those, whose survival status was not correctly predicted