Background and Problem: The loss of a kidney graft represents nowadays an important multifactorial cause of end-stage renal disease worldwide. Current prediction systems of graft loss do not integrate the dynamic effect of parameters assessed over the time course of kidney recipients.
Solution: Important algorithmic developments have been made in the past years and allow the integration of repeated measurements in predictive survival analysis. These dynamic approaches could permit to construct a system continuously updatable over time thereby improving patient care and treatment management.
Methods: Joint modelling was used to derive a dynamic prediction model. This method associates 1) a survival model for parameters measured at one timepoint, with 2) combined mixed models to integrate repeated measurements.
We applied this approach to an international cohort study involving 21 transplant centers and 7 randomized controlled trials (RCT). Patients were divided into a derivation cohort (in France) and validation cohorts (in Europe, US, South America). Patients underwent assessment of clinical, functional, histological and immunological parameters (gathered in the iBox score previously published by our team1), together with glomerular filtration rate (GFR) and proteinuria repeated measurements (gathered in combined mixed models). The outcome was the graft loss.
Results: 15,028 patients were included (3,774 patients in the derivation cohort, 9,085 in the validation cohorts and 2,169 in the RCTs). After a median follow-up of 7.4 years post transplantation, 1,408 graft failures occurred. 416,510 GFR and proteinuria repeated measurements were assessed. With joint modeling, the iBox score and the repeated measurements were independently associated with graft loss. Based on the final multivariable model, we derived a dynamic risk prediction model that demonstrated accurate calibration and very high discrimination in the derivation cohort (AUC = 0.857). The performance of the model was confirmed in the four validation cohorts from Europe (AUC = 0.833), the USA (AUC = 0.897), South-America (AUC = 0.891), and the RCTs (AUC = 0.922).
We developed for the first time an integrative dynamic system that accurately predicts the risk of long-term graft failure and outperforms any current prediction models in kidney transplantation based on classical statistical approaches. This dynamic system shows generalisability across centers and countries worldwide and may help adjusting prognostic judgements of clinicians in everyday practice and improve the design of future clinical trials (Trial registration number: NCT03474003).
References: 1Loupy A, Aubert O, Orandi BJ, et al. Prediction system for risk of allograft loss in patients receiving kidney transplants: international derivation and validation study. BMJ (2019;366:l4923)