Perioperative prediction of ideal postoperative recovery after kidney transplantation: a multicenter retrospective cohort study

Abstract

Kidney transplantation is the preferred treatment for end-stage renal disease, but early postoperative outcomes remain heterogeneous. Textbook outcome (TO), a composite measure of an ideal postoperative recovery trajectory, has emerged as a meaningful indicator of surgical quality; however, robust perioperative prediction of TO remains insufficiently developed, particularly across multiple centers. This retrospective multicenter study screened 1141 kidney transplant recipients treated between 2016 and 2023 and included 880 patients with complete perioperative clinical data and preoperative donor-kidney noncontrast computed tomography. One center served as the training cohort ( n = 357), and 2 centers served as independent external validation cohorts ( n = 302 and n = 221). Donor–recipient clinical variables, including HLA mismatch, were collected. After three-dimensional parenchymal segmentation, 1316 radiomic features were extracted from donor-kidney images. Clinical-only, radiomic-only, and combined clinical-radiomic models were developed using 9 machine-learning algorithms. Overall, 432 of 880 recipients (49.1%) achieved TO. Among the nine algorithms, the combined CatBoost model achieved AUCs of 0.919 (95% CI 0.887–0.951), 0.817 (95% CI 0.736–0.898), and 0.803 (95% CI 0.718–0.887) in the training cohort and the two external validation cohorts, respectively, and provided the most favorable overall balance of discrimination, calibration, and clinical net benefit. Important contributors included ureteral anastomosis type, HLA mismatch, donor hypertension, donor intensive care unit stay, donor serum creatinine, and cold ischemia time. An interpretable CatBoost model integrating perioperative clinical variables and preoperative donor-kidney radiomic features predicts TO after kidney transplantation and may support perioperative risk stratification and clinical planning.