Abstract
Objective This study aimed to develop a machine learning-based model for the early prediction of multi-drug resistant organism(MDRO)infection in patients with decompensated cirrhosis,providing a clinical decision-making tool for the early identification of high-risk groups.Methods A total of 220 patients with decompensated cirrhosis,hospitalized at Suqian First Hospital from March 2022 to November 2024,were retrospectively enrolled.Based on the occurrence of MDRO infection,they were classified into an MDRO infection group(n=61)and a non-MDRO infection group(n=159).Logistic regression analysis was performed to identify characteristic variables for the prediction model.The patients were randomly partitioned into a training set(154 cases)and a test set(66 cases)in a 7:3 ratio.Five predictive models—Support Vector Machine(SVM),Logistic Regression(LR),Gradient Boosting Decision Tree(GBDT),Random Forest(RF),and Extreme Gradient Boosting(XGBoost)—were constructed using machine learning algorithms.The predictive performance of each model was evaluated and compared.The optimal model was interpreted using SHapley Additive exPlanations(SHAP)values.A total of 68 patients with decompensated liver cirrhosis who met the same inclusion and exclusion criteria,admitted to Suqian First Hospital between December 2024 and March 2025,were selected as the validation cohort for external validation of the optimal model.Results Among 220 patients with decompensated cirrhosis,the incidence of MDRO infection was 27.73%(61/220).A total of 89 strains were isolated from these 61 patients,predominantly Gram-positive bacteria,with Enterococcus faecium(48.32%)and Enterococcus faecalis(16.85%)constituting 65.17%of the isolates.Gram-negative bacteria accounted for 33.71%of the isolates.Analysis revealed statistically significant difference between the two groups in terms of Model for End-stage Liver Disease(MELD)score,length of hospital stay,invasive procedures,hypoalbuminemia,antibiotic use,and C-reactive protein-to-albumin ratio(CAR)(P<0.05).Multivariate Logistic regression identified all these six indicators as independent risk factors for MDRO infection(P<0.05).Among the five prediction models—SVM,LR,GBDT,RF,and XGBoost—the XGBoost model achieved superior performance on the test set,demonstrating the highest area under the curve(AUC),accuracy,recall,precision,and F1-score.SHAP visualization analysis indicated that all feature variables in the XGBoost model significantly influenced MDRO infection,with feature importance ranked as follows:CAR,length of hospital stay,MELD score,hypoalbuminemia,invasive procedures,and antibiotic use.Analysis of a validation set showed that the XGBoost model had an accuracy of 88.24%and a Kappa value of 0.722(95%CI 0.613 to 0.850)for predicting MDRO infection.Conclusion The XGBoost prediction model developed using the MELD score,length of hospital stay,invasive procedures,hypoalbuminemia,antibiotic use,and CAR provides a quantitative tool for the early identification,risk stratification,and clinical intervention of MDRO infection in patients with decompensated cirrhosis,thereby improving patients prognosis and promoting the rational application of antibacterial agents.关键词
肝硬化/失代偿期/多重耐药菌感染/机器学习/XGBoost模型/风险预测/SHAP值Key words
cirrhosis/decompensated stage/multi-drug resistant organism(MDRO)infection/machine learning/XGBoost model/risk prediction/SHAP value分类
医药卫生
