Predicting Multidrug-Resistant Pneumonia: An Interpretable Machine Learning Model Validated in US and Chinese Patient Cohorts.

Objective: Multidrug-resistant organisms (MDROs) complicate hospital-acquired and ventilator-associated pneumonia (HAP/VAP). We aimed to develop and validate a machine learning (ML) model to predict MDR risk in pneumonia patients and assess its utility for clinical decision support. Patients and Methods: We developed multiple ML models using data from the MIMIC-IV database (n=802). Feature selection was performed using LASSO regression and chi-square tests. Four models-Logistic Regression, Random Forest, XGBoost, and LightGBM-were trained, tuned, and calibrated. Model performance was evaluated using ROC curves and calibration plots. The final model was selected based on discrimination, calibration, and interpretability (assessed via SHAP). External validation on an independent cohort from a Chinese tertiary hospital (n=213) demonstrated the reproducibility and generalizability of its performance. Results: Among the evaluated ML models, Logistic Regression demonstrated the best overall performance. On the MIMIC-IV internal test set, it achieved an area under the curve (AUC) of 0.798 (95% CI: 0.718-0.872) with an accuracy of 0.807. External validation on an independent Chinese cohort confirmed the model's robust generalizability, achieving an AUC of 0.845. SHAP analysis identified key predictive features consistently across both cohorts, including the systemic immune-inflammation index (SII), albumin level, C-reactive protein-to-albumin ratio (CAR), number of antibiotic classes, white blood cell count (WBC), and absolute lymphocyte count (LYM abs). All of these features were significantly associated with MDR risk. Conclusion: Multiple ML models effectively predicted MDR infections in pneumonia patients, with Logistic Regression exhibiting particularly strong overall performance. Model reliability was enhanced through feature selection and probability calibration, while interpretability was improved by SHAP analysis. External validation confirmed the generalizability of our approach, supporting its potential application in clinical infection control. Future studies should focus on validation and the integration of more diverse clinical data sources.