Predicting the graft survival for kidney transplantation is a high stakes undertaking considering the shortage of available organs and the utilization of healthcare resources. The strength of any predictive model depends on the selection of proper predictors. However, despite improvements in acute rejection management and short-term graft survival, the accurate prediction of kidney transplant outcomes remains suboptimal. Among other approaches, machine-learning techniques have the potential to offer solutions to this prediction problem in kidney transplantation. This study offers a novel methodological solution to this prediction problem by: (a) analyzing the retrospective database including > 31,000 U.S. patients; (b) introducing a comprehensive feature selection framework that accounts for medical literature, data analytics methods and elastic net (EN) regression (c) using sensitivity analyses and information fusion to evaluate and combine features from several machine learning approaches (i.e., support vector machines (SVM), artificial neural networks (ANN), and Bootstrap Forest (BF)); (d) constructing several different scenarios by merging different sets of features that are optioned through these fused data mining models and statistical models in addition to expert knowledge; and (e) using best performing sets in Bayesian belief network (BBN) algorithm to identify non-linear relationships and the interactions between explanatory factors and risk levels for kidney graft survival. The results showed that the predictor set obtained through fused data mining model and literature review outperformed the all other alternative predictors sets with the scores of 0.602, 0.684, 0.495 for F-Measure, Average Accuracy, and G-Mean, respectively. Overall, our findings provide novel insights about risk prediction that could potentially help in improving the outcome of kidney transplants. This methodology can also be applied to other similar transplant data sets.