Pulled from PubMed
Francesco DI Bello 1 2, Andrea Gallioli 3, Alessio Pecoraro 4, Thomas Prudhomme 5, Alberto Piana 6, Beatriz Bañuelos Marco 7, Hakan B Haberal 8, Muhammet I Dönmez 9, Alicia Lopez-Abad 10, Donato Cannoletta 3 11, Stefano Mancon 3 12, Lluis Guirado 13, Carma Facundo 13, Andres K Kanashiro 3, Pavel Gavrilov 3, Oscar Rodriguez-Faba 3, Josep M Gaya 3, Alberto Breda 3, Angelo Territo 3; EAU – YAU Kidney transplant working group
Introduction: Gradient-boosting (GB) algorithm is considered as the state-of-the-art algorithm for prediction of survival. The aim of the current study was consolidating the evidence on GB machine-learning (ML) model to predict graft survival (GS) after kidney transplant (KT).
Evidence acquisition: A systematic search (PROSPERO: CRD42025645353) with a qualitative analysis was performed according to PRISMA statement. Study quality and risk of bias were evaluated using the Prediction-model Risk of Bias ASsessment Tool (PROBAST).
Evidence synthesis: Overall, 15 studies involving 889,657 patients were included in the final analysis. Of those, 14,334 included GS information. According to ML algorithm, 12 (80.3%) studies relied on eXtremeGB, two (13.3%) on StochasticGB and one (6.4%) on lightGB. The model performance was evaluated with Area Under Curve (AUC) methodology in 12 (80%) of papers and ranged from 0.715 to 0.989. The Brier-score was evaluated in five (33.3%) papers and ranged from 0.020 to 0.14. The C-index and/or Accuracy were evaluated in three (20%) papers and ranged, in respectively, from 0.635 to 0.837, and from 0.81 to 0.979.
Conclusions: The current systematic review showed a promising potential role of GB in the GS prediction after KT. However, ML models should be carefully interpreted before being used in clinical practice.