Introduction: Prediction models of graft survival in kidney transplantation allow risk stratification of kidney transplant recipients using big data and artificial intelligence. Many models require detailed information on allograft histology and presence of donor-specific antibodies which are not readily available in every transplant center. We developed a joint model using Bayesian estimation to correlate baseline transplant characteristics and longitudinal measurements of both serum creatinine (S-Cr) and urine protein-creatinine ratios (u-PCR) with death-censored graft failure (DCGF).

Methods: 3596 kidney transplant recipients (KTRs) from four transplant centers in Europe were included with a total of 452.033 S-Cr and u-PCR measurements. Each KTR had a median of 112 (IQR: 52-163) consecutive measurements. The interval between measurements was 7 (IQR: 1-35) days. Median follow-up was 6 (IQR: 3-9) years. Model predictors included recipient and donor gender and age, total number of human leukocyte antigen-mismatches, donor type, pre-transplant DSAs, and dialysis vintage. Primary outcomes were DCGF six months after last S-Cr or u-PCR measurement and DCGF seven years post-transplantation. Two training and external validation methods have been used to assess best model performance: model 1; development in one center and validation in three remaining centers, and model 2; development on 80% of the total cohort and validation on the remaining 20% (Table 1).

Results: A total of 549 (15%) out of 3596 KTRs experienced DCGF. Joint modeling revealed that recipient age, recipient gender, donor age, transplantation after dialysis initiation, and serial measurements of S-Cr and u-PCR were independent risk factors for DCGF. Our final joint models showed good calibration (prediction error: 0.01) and very high discrimination in the development and validation cohorts (6-months incident AUC 0.9 [95% CI: 0.87-0.93] and 7-years dynamic AUC 0.84 [95% CI: 0.81-0.87] for model 1, and 6-months incident AUC 0.98 [95% CI: 0.96-1.00] and 7-years dynamic AUC 0.85 [95% CI: 0.80-0.91] for model 2) (Table 1). Especially model 2 could be of interest as a personalized surveillance tool in clinical practice given its AUCs >0.95 from year 1 onwards.

Conclusion: We developed a high performing and multicenter-validated, dynamic prediction model on death-censored allograft survival based on traditionally available baseline transplant characteristics and longitudinal measurements of S-Cr and u-PCR. This dynamic model can continuously be updated with new measurements and might enable personalized surveillance of kidney transplant recipients and objectify allograft prognosis (as illustrated in Figure 1). Of notice, our models can be extended with newly discovered dynamic biomarkers in the future if needed.