*Purpose: Identification of patients at risk for graft failure can improve individual surveillance and early therapeutic interventions. We used a well-documented patient cohort to establish a prediction model for graft survival and to identify and assess the risk factors for graft loss.

*Methods: A cohort of 892 patients was used to build a model that predicts 15-year graft survival censored for death, applying a machine learning-based classification algorithm for right censored data. Pre-transplant and first-year posttransplant data were included, along with results from protocol biopsies at 6 weeks, 3 and 6 months and additional biopsies for cause. A cohort of 349 patients was used to validate the model. Loss to follow-up was negligible (none in the validation cohort, n=15 in the training cohort).

*Results: Patient and graft survival over 15 years was 50% and death-censored graft survival was 71% in the training cohort. Major causes of graft loss were acute and chronic rejection (19%), other specified histomorphological lesions (14%), and progressive chronic dysfunction without histomorphological specification (34%). The final model predicted graft survival with a concordance index of 0.79, with stable predictive performance over the entire posttransplant time course. Application of the model to the validation cohort resulted in a concordance index of 0.78. Besides the GFR at 12 months after transplantation, main predictive factors in the model were T cell- and antibody-mediated rejection (with higher impact of rejections in biopsies for cause than in protocol biopsies), recurrent and de novo renal disease, BK virus nephritis, overweight, diabetes, coronary heart disease, monoclonal gammopathy, replicative hepatitis B and other infections, and donor factors. Using cut-points of the prognostic index at the 16^th, 50^th and 84^th percentiles, calibration analyses showed good discrimination of risks for graft loss by the model, with 15-year graft survival of 100% in patients with very low risk, of 93% with low risk, of 50% with moderate and 18% with high risk.

*Conclusions: The established model allows reliable assessment of the individual risk for graft loss and the associated factors. Based on the model, patients can be stratified in different algorithms for monitoring and treatment decisions.

« Back to 2020 American Transplant Congress