Abstract
BACKGROUND The optimization of tacrolimus dosing during the early postoperative hospitalization period is essential to prevent rejection, minimize nephrotoxicity, and minimize the risk of opportunistic infections. Patient pharmacokinetic variability poses challenges in dose adjustment. This study aimed to evaluate tacrolimus dosing optimization using machine learning and statistical methods. MATERIAL AND METHODS We conducted a retrospective study of 749 kidney transplant recipients at Seoul St. Mary's Hospital between January 2015 and December 2019. Data on tacrolimus doses, trough levels, and other clinical variables were collected and analyzed during the first 12 postoperative days of hospitalization. Three approaches were evaluated: Extreme Gradient Boosting (XGBoost), Elastic Net regression (EN), and Linear regression (LR). The models were trained and validated using 5-fold cross-validation, with performance assessed using R² errors and alignment with clinically acceptable error margins. RESULTS Elastic Net showed the best performance with R² (Coefficient of Determination) of 0.861±0.044 and RMSE (Root Mean Square Error) of 0.930±0.220. Linear Regression and XGBoost provided clinically relevant predictions but with slightly lower accuracy. External validation was not performed, limiting the generalizability of the results. CONCLUSIONS The Elastic Net is a practical and reliable model for predicting the optimal tacrolimus dose. Machine learning and statistical methods are useful tools for optimizing tacrolimus dosing during hospitalization after kidney transplantation. Future studies should incorporate multi-center validation to improve clinical applicability.