Establishment and verification of risk prediction models for postoperative multidrug-resistant organisms infections in liver transplantation patients based on 7 types of machine learning algorithm

OBJECTIVE To establish and verify the risk prediction models for postoperative multidrug-resistant organisms (MDROs) infections in the liver transplantation patients based on the machine learning algorithms so as to provide bases for identification of the population at high risk of postoperative MDROs infections.

METHODS The liver transplantation patients who were retrospectively collected from intensive care Ⅳ database (MIMIC-Ⅳ) and eICU collaborative research database (eICU)were recruited as the research subjects, meanwhile, the patients who underwent liver transplantation in the First Hospital of Shanxi Medical University from Jan. 2021 to Jul. 2024 were assigned as the external verification group. The variables were selected by Lasso regression, and the models were established based on 7 types of machine learning algorithms such as extreme gradient boosting algorithm and random forest. The predictive performances of the models were evaluated by comparing the areas under receiver operating characteristic (ROC) curves and the accuracy, the characteristic variables were interpreted by Shapley additive explanations (SHAP), and the risk prediction calculator was established.

RESULTS A total of 637 patients were finally enrolled in the study, and the incidence of postoperative MDROs infections was 35.79%. Totally 15 variables were finally selected for construction of the model. The area under the receiver operating characteristic curve of XGBoost model was 0.82 for the internal test set, 0.78 for the external test set; the predictive performance of XGBoost model was better than that of the rest of 6 models. SHAP algorithm indicated that the top 5 important predictive factors were as follows: hepatic encephalopathy, length of intensive care unit (ICU) stay, albumin, model of end-stage liver disease (MELD) and total length of hospital stay.

CONCLUSION The risk prediction models that are established based on the machine learning algorithms have remarkable effect on prediction of the postoperative MDROs infections and can accurately identify the liver transplantation patients at high risk of postoperative MDROs infections, which may provide guidance for the identification of high-risk population and the development of prevention and treatment measures for infections.