*Purpose: Delayed graft function (DGF) has been associated with adverse kidney transplant outcomes. Accurate prediction of DGF before surgery will help select organs and candidates to improve post-transplant outcomes. Previously developed prediction models using multivariable logistic regression, however, have limitations. Since machine learning (ML) has the potential to exploit previously underutilized data, we aimed to evaluate ML models to predict DGF after deceased donor kidney transplantation with sequential data.

*Methods: We analyzed single-center data of deceased-donor single-kidney transplants from 2010 to 2018 (n=1797). DGF was defined as the need for dialysis within seven days of transplant surgery. Using a priori donor and recipient variables, we developed models to predict DGF. In addition to donor and recipient demographic data, we utilized sequential donor data (i.e., changes in urine output within 24 hours before organ recovery and serum creatinine levels during donor hospitalization) to test the values in prediction. We trained the model using 70% of the data, tuned the parameters using grid search cross-validation, and tested the model using the remaining 30%. We also evaluated the role of soft voting (majority rule) of different models in classification.

*Results: When the only static donor and recipient features were used, the conventional logistic regression model showed relatively low balanced accuracy (BA): 0.6375, F1 score: 0.4522, and C-statistic (ROC AUC): 0.7532. Different ML models using the same static dataset showed various prediction scores (Table 1). Soft voting of the ML models (catboost, extra trees, random forest, and linear discriminant analysis) demonstrated slightly increased BA: 0.6855, F1: 0.5509, and ROC AUC: 0.7628. Although sequential donor data (i.e., urine volume and serum creatinine) alone had low predictive values, the addition of such sequential data to the static donor and recipient characteristics further increased prediction scores (BA: 0.6912, F1: 0.5566, and ROC AUC: 0.7702).

*Conclusions: The soft-voting ML models provided better DGF prediction scores with conventional donor and recipient characteristics. The ML models further improved prediction scores by exploiting the previously underutilized sequential data. Further studies are warranted to validate and generalize this strategy in other cohorts and clinical settings.

« Back to 2022 American Transplant Congress