Machine Learning-Based Prediction Model for Multidrug-Resistant Organisms Infections: Performance Evaluation and Interpretability Analysis

Wenting Zhao,1,2,* Pei Sun,1,2,* Wei Li,3 Linping Shang2,4 1College of Nursing, Changzhi Medical College, Changzhi, Shanxi, People’s Republic of China; 2College of Nursing, Shanxi Medical University, Taiyuan, Shanxi, People’s Republic of China; 3Infection Management Department, the First Hospital of Shanxi Medical University, Taiyuan, Shanxi, People’s Republic of China; 4Nursing Department, the First Hospital of Shanxi Medical University, Taiyuan, Shanxi, People’s Republic of China*These authors contributed equally to this workCorrespondence: Linping Shang, College of Nursing, Shanxi Medical University, Taiyuan, Shanxi, People’s Republic of China, Email [email protected]: Multidrug-resistant organism (MDRO) infections pose a significant global health threat, particularly in intensive care units (ICUs), where delayed identification exacerbates clinical outcomes. Although machine learning (ML) holds promise for infection prediction, the opaque nature of complex algorithms impedes clinical adoption. This study evaluated an interpretable machine learning model incorporating SHapley Additive exPlanations (SHAP) to predict MDRO infections in ICU patients.Methods: A retrospective cohort study was conducted on 888 ICU patients (2020– 2022) from a tertiary hospital in China. Following TRIPOD guidelines, key predictors were identified using Lasso regression from a comprehensive set of clinical variables, including demographics, treatments, and laboratory data. Six machine learning algorithms—Neural Networks, Random Forests, Support Vector Machines, Logistic Regression, Decision Trees, and Gaussian Naive Bayes—were evaluated based on AUC, accuracy, and calibration curves. SHAP analysis provided both global and local interpretability.Results: Among 825 eligible cases (375 MDRO infections), the Random Forest model exhibited the highest performance (AUC = 0.83, accuracy = 76.7%). SHAP analysis identified urinary catheterization, ventilator use, and prolonged antibiotic exposure as key modifiable risk factors. Case-level interpretation via dynamic force plots illustrated individualized risk stratification. Decision curve analysis indicated clinical utility within probability thresholds of 0.44– 0.60.Conclusion: This study establishes an interpretable prediction framework integrating RF algorithms with SHAP explainability, balancing predictive accuracy with clinical transparency. The model’s dynamic visualization capabilities support individualized risk assessment and evidence-based antimicrobial stewardship. Integration into hospital information systems with real-time dashboards could enhance early intervention strategies.Keywords: MDRO, machine learning, prediction, intensive care unit