A model-driven machine learning approach for personalized kidney graft risk prediction

Abstract

Graft failure after renal transplantation is a multifactorial process. Predicting the risk of graft failure accurately is imperative since such knowledge allows for identifying patients at risk and treatment personalization. In this study, we were interested in predicting the temporal evolution of graft function (expressed as estimated glomerular filtration rate; eGFR) based on pretransplant data and early post-operative graft function. Toward this aim, we developed a tailored approach that combines a dynamic GFR mathematical model and machine learning while taking into account the corresponding parameter uncertainty. A cohort of 892 patients was used to train the algorithm and a cohort of 847 patients for validation. Our analysis indicates that an eGFR threshold exists that allows for classifying high-risk patients. Using minimal inputs, our approach predicted the graft outcome with an accuracy greater than 80% for the first and second years after kidney transplantation and risk predictions were robust over time.