Machine learning prediction of 30-day bleeding complications in desensitized kidney transplant recipients.

초록

Background: Desensitization protocols enable kidney transplantation in immunologically high-risk recipients but may increase hemorrhagic complications through altered coagulation and immune pathways. Accurate preoperative risk stratification could guide perioperative management, yet no validated prediction tool exists for this population. Methods: We analyzed kidney transplant recipients recorded in the Korean Quality Improvement Platform in Surgery–Kidney Transplant (KQIPS-KT) registry, a prospective nationwide multicenter registry covering more than 70% of kidney transplants in Korea. Among 14,487 recipients from 52 centers, 2,802 recipients at 25 desensitization-capable centers who underwent preoperative desensitization comprised the prediction cohort. The primary outcome was 30-day bleeding complication, defined as intraoperative transfusion or postoperative hemorrhage causing hemodynamic instability requiring transfusion, or CT-confirmed bleeding requiring intervention or reoperation. We first examined the association between desensitization and bleeding using multivariable logistic regression with multiple imputation and inverse probability of treatment weighting (IPTW). We then developed prediction models among desensitized recipients using 87 preoperative covariates, comparing LASSO logistic regression, random forest, XGBoost, and TabPFN (a pre-trained tabular foundation model). The cohort was split into training (80%) and held-out test (20%) sets; model selection was performed by 5-fold stratified cross-validation within the training set, with the test set reserved exclusively for final evaluation. The winning algorithm was then optimized into a deployable parsimonious model through feature count optimization (Kneedle elbow method with pre-specified selection rules) and calibration assessment, with final evaluation on the held-out test set. Results: The 30-day bleeding complication rate was 27.2% (763/2,802) among desensitized recipients versus 14.2% (1,620/11,419) among non-desensitized recipients (IPTW adjusted OR 1.48, 95% CI 1.32–1.66, p<0.001). TabPFN outperformed LASSO, random forest, and XGBoost, and was optimized into a parsimonious 18-feature model. On the held-out test set (n=561, 153 events), this model achieved AUROC 0.83 (95% CI 0.80–0.87), AUPRC 0.67 (0.59–0.73), and Brier score 0.142 with adequate calibration (Hosmer-Lemeshow p=0.332; Figure 1). Performance was robust across sex, age, donor type, rituximab dose, plasmapheresis sessions, and re-transplant status (AUROC range 0.71–0.91). SHAP analysis identified preoperative hematocrit as the dominant predictor, followed by preoperative length of stay, antiplatelet use, donor type, and CRP (Figure 2a). The model offers flexible operating points from high-sensitivity screening (93.5% sensitivity, 95.4% NPV at threshold 0.15) to high-specificity confirmation (81.6% specificity, 55.6% PPV at threshold 0.35). The model stratified patients into low (<15% predicted probability; 38.9% of cohort, observed rate 4.6%), medium (15–30%; 25.7%, rate 25.0%), and high (>30%; 35.5%, rate 53.8%) risk groups — an 11.7-fold gradient (Figure 2b). Conclusion: Desensitized kidney transplant recipients had substantially higher bleeding risk than non-desensitized recipients. A TabPFN-based model using routinely available preoperative data provided accurate prediction and clinically meaningful risk stratification for 30-day bleeding complications.