Prediction of delayed graft function after kidney transplantation using artificial intelligence and machine learning:...

Delayed graft function (DGF) is a frequent early complication after kidney transplantation. We systematically reviewed artificial intelligence and machine learning (AI/ML) models that use routine clinical variables to predict DGF.

This PROSPERO-registered review (CRD420261400263) followed PRISMA guidance. MEDLINE/PubMed, Embase, Web of Science Core Collection and Cochrane Library were searched for human studies, without date restriction. Studies that developed or validated AI/ML models for DGF prediction using routine donor, recipient, immunological, laboratory, ischaemia-time, transplant-process or perioperative variables were eligible. Results were synthesised narratively. Quality was assessed using PROBAST+AI and STREAM-URO.

Of 613 records, 12 studies were included. Most were retrospective deceased-donor cohorts and three used national registries. DGF incidence varied widely (12.5-56.8%). Algorithms included neural networks, logistic and penalised regression, support vector machines, random forests, gradient boosting and stacked ensembles. Reported areas under the receiver operating characteristic curve (AUCs) ranged from 0.67 to 0.99. Recurrent predictors included donor renal-function measures, donor age and body mass index, donation after circulatory death, ischaemia time, recipient body size, dialysis exposure and human leukocyte antigen matching. Calibration, decision-curve analysis and external validation were uncommon. PROBAST+AI indicated high overall risk of bias in eight studies.

AI/ML models show promise for DGF prediction, but their incremental benefit over conventional regression remains inconsistent. Heterogeneity across transplant eras, limited model transparency, inadequate assessment of calibration and clinical utility, and the lack of robust external validation must be addressed before these models can be implemented routinely in clinical practice.