Objective  To compare multiple models and analyze SHAP feature contributions to construct a machine learning risk prediction model for post-reperfusion syndrome (PRS) during adult living donor liver transplantation (LDLT).

Methods  Clinical data of 390 patients who underwent LDLT due to end-stage liver disease at the First Affiliated Hospital of Anhui Medical University from May 2023 to April 2025 were collected. Five machine learning models, including random forest, logistic regression, XGBoost, decision tree, and AdaBoost, were compared. Feature selection was performed using Lasso regression, and the training set was divided into five folds for stratified cross-validation. The generalization ability of the model was evaluated on the independent test set. The models were comprehensively compared based on key evaluation indicators such as recall rate, accuracy rate, precision rate, area under the curve (AUC) and F1 value, to determine the optimal model.

Results  Eight potential factors for predicting PRS were selected. The random forest model demonstrated the best prediction performance in both the training set and the test set. Its accuracy rate was as high as 84.2% (AUC = 0.894, 95% confidence interval 0.808-0.964) in the test set. The importance ranking of PRS predictors was determined through SHAP value analysis. Cold ischemia time, K⁺ during the anhepatic period and the model for end-stage liver disease (MELD) score were all identified as the most significant predictors.

Conclusions  Based on eight key indicators including cold ischemia time, portal vein occlusion time, pre-reperfusion body temperature, alkaline reserve during the anhepatic period, K⁺ during the anhepatic period, MELD score, left ventricular end-diastolic diameter and graft volume to standard liver volume ratio, the PRS prediction model constructed using the random forest algorithm demonstrates the best prediction performance in the test set, providing certain assistance for subsequent clinical predictions.