Abstract
Background: Kidney allograft failure is a common cause of end-stage renal disease. We aimed to develop a dynamic artificial intelligence approach to enhance risk stratification for kidney transplant recipients by generating continuously refined predictions of survival using updates of clinical data.Methods: In this observational study, we used data from adult recipients of kidney transplants from 18 academic transplant centres in Europe, the USA, and South America, and a cohort of patients from six randomised controlled trials. The development cohort comprised patients from four centres in France, with all other patients included in external validation cohorts. To build deeply phenotyped cohorts of transplant recipients, the following data were collected in the development cohort: clinical, histological, immunological variables, and repeated measurements of estimated glomerular filtration rate (eGFR) and proteinuria (measured using the proteinuria to creatininuria ratio). To develop a dynamic prediction system based on these clinical assessments and repeated measurements, we used a Bayesian joint models-an artificial intelligence approach. The prediction performances of the model were assessed via discrimination, through calculation of the area under the receiver operator curve (AUC), and calibration. This study is registered with ClinicalTrials.gov, NCT04258891.Findings: 13 608 patients were included (3774 in the development cohort and 9834 in the external validation cohorts) and contributed 89 328 patient-years of data, and 416 510 eGFR and proteinuria measurements. Bayesian joint models showed that recipient immunological profile, allograft interstitial fibrosis and tubular atrophy, allograft inflammation, and repeated measurements of eGFR and proteinuria were independent risk factors for allograft survival. The final model showed accurate calibration and very high discrimination in the development cohort (overall dynamic AUC 0·857 [95% CI 0·847-0·866]) with a persistent improvement in AUCs for each new repeated measurement (from 0·780 [0·768-0·794] to 0·926 [0·917-0·932]; p<0·0001). The predictive performance was confirmed in the external validation cohorts from Europe (overall AUC 0·845 [0·837-0·854]), the USA (overall AUC 0·820 [0·808-0·831]), South America (overall AUC 0·868 [0·856-0·880]), and the cohort of patients from randomised controlled trials (overall AUC 0·857 [0·840-0·875]).Interpretation: Because of its dynamic design, this model can be continuously updated and holds value as a bedside tool that could refine the prognostic judgements of clinicians in everyday practice, hence enhancing precision medicine in the transplant setting.