OBJECTIVE  To explore the risk factors for infection with multidrug-resistant organisms （MDRO） in patients with severe pneumonia in the intensive care unit （ICU）, and to construct a predictive model for MDRO infection based on machine learning, thereby providing a reference for clinical decision-making.

METHODS  Clinical data from 202 ICU patients with severe pneumonia admitted to Hongqi Hospital Affiliated to Mudanjiang Medical University from Jan. 2021 to Dec. 2023 were retrospectively collected to analyze the infectious pathogens and their drug resistance. Patients were divided into an MDRO group （n=67） and a non-MDRO group （n=135） according to the presence or absence of MDRO infection after ICU admission. Logistic regression analysis was employed to identify risk factors for MDRO infection. With the occurrence of MDRO infection as the final outcome, the dataset was randomly divided into a training set （70%） and a test set （30%）. Three machine learning models （Random Forest, Gradient Boosting and AdaBoost） were developed. Each model underwent five-fold cross-validation, and the area under the curve （AUC） and positive predictive rate were adopted as the primary indicators to evaluate model performance.

RESULTS  A total of 271 strains of pathogenic bacteria were isolated from the 202 patients with severe pneumonia, including 193 strains （71.22%） of gram-negative bacteria （mainly Klebsiella pneumoniae）, 45 strains （16.60%） of fungi and 33 strains （12.17%） of gram-positive bacteria. The isolation rate of MDROs was 19.19%（52/271）, and the isolation rate of extended-spectrum β-lactamases （ESBLs）-producing in K. pneumoniae was up to 30.56%（22/72）. The length of ICU stay （OR=1.092） and the APACHE II score （OR=1.123） were identified as risk factors for MDRO infection in ICU patients with severe pneumonia. Among the three machine learning models, the Gradient Boosting algorithm achieved the highest AUC （0.916）. Although the AUC values of the Random Forest and AdaBoost algorithms were also ＞0.9, their positive predictive rates were lower than that of the Gradient Boosting algorithm. Therefore, the Gradient Boosting algorithm model demonstrated superior predictive capability.

CONCLUSIONS  The Gradient Boosting model exhibits the best performance, identifying predictive factors such as an elevated neutrophil ratio, prolonged mechanical ventilation duration, abnormal platelet count and the use of ≥3 antimicrobial agents. This provides a reference for early warning and individualized prevention and control of MDRO infection.