Leveraging Machine Learning to Predict Delayed Graft Function Occurrence and Length in Kidney Transplant Recipients

Predicting delayed graft function (DGF) relies on donor, recipient, and perioperative factors. Despite the growing recognition that DGF duration strongly influences patient outcomes, no models currently address or predict its impact-highlighting an important gap in current research and clinical practice. This study aimed to develop an ensemble-based machine learning model using perioperative data to predict DGF occurrence and duration. The gradient-boosted decision trees model was trained and validated on 2725 patients, with k-fold cross-validation in an external cohort of 284 patients. Model performance was evaluated based on accuracy, ROC-AUC, and other metrics using R and Python libraries. The binary DGF prediction model achieved a ROC-AUC of 0.77, while the DGF duration classification model had an accuracy of 79.2%. DGF duration AUC values by time interval were 0.76 for 0-1 weeks, 0.81 for >1-2 weeks, 0.80 for >2-3 weeks, and 0.87 for >3 weeks; the overall macro-averaged AUC was 0.81. The mean Brier score for multi-class predictions was 0.14. External validation showed a 78% accuracy for DGF duration prediction. Acute kidney injury (AKI) and donor donation after circulatory death (DCD) status were key DGF predictors. The use of gradient-boosted decision trees (GBDT) improves the prediction of both the likelihood and duration of DGF, addressing a current gap in kidney transplant patient care. By facilitating personalized transplant care, this model supports more effective perioperative planning and timely interventions, which may contribute to better patient outcomes.