Construction and internal-external validation of a machine learning-based risk prediction model for multidrug resistance in ICU patients with acute exacerbation of chronic obstructive pulmonary disease.

OBJECTIVE

This study aimed to create and validate a machine learning (ML) model for predicting the risk of multidrug-resistant (MDR) infection in critically ill patients with acute exacerbation of chronic obstructive pulmonary disease (AECOPD) admitted to the intensive care unit (ICU).

METHODS

Data from patients diagnosed with AECOPD were retrospectively extracted from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database. A total of 1,018 patients were split at a 7:3 ratio into a training set (n = 712) for model development and an internal validation set (n = 306). MDR occurrence served as the outcome. Features were selected with the Boruta algorithm and least absolute shrinkage and selection operator (LASSO). Five ML algorithms-XGBoost, support vector machine (SVM), LightGBM, CatBoost, and Gaussian naïve Bayes (NB)-were trained with 10-fold cross-validation. Model performance was evaluated by the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1 score, calibration plot, decision curve, and clinical impact curve. An external validation cohort of 324 AECOPD patients admitted to the Affiliated Hospital of Jiangsu University between January 2024 and October 2025 was collected. The best-performing model was interpreted with Shapley Additive exPlanations (SHAP) to clarify feature importance and decision logic, and a nomogram was constructed to increase readability.

RESULTS

Boruta identified SpO₂, RBC, hemoglobin, hematocrit, blood urea nitrogen, creatinine, Nbpd, and Nbpm as significant predictors. The LightGBM outperformed the other algorithms: in the internal validation set, it achieved 78.9% accuracy, 90.5% sensitivity, 87.3% specificity, 60.5% F1 score, and an AUC of 0.966 (95% CI 0.951-0.981); in the external validation set, it reached 74.4% accuracy, 71.1% sensitivity, 77.8% specificity, 58.7% F1 score, and an AUC of 0.926 (95% CI 0.895-0.958). SHAP analysis indicated that hematocrit and SpO₂ were the primary drivers of model decisions. Interactive and dynamic nomograms were successfully developed.

CONCLUSION

Multidrug-resistant occurrence in AECOPD patients was associated with SpO₂, RBC, hemoglobin, hematocrit, blood urea nitrogen, creatinine, Nbpd, and Nbpm. The LightGBM model demonstrated good discriminative ability but limited sensitivity for detecting positive cases, offering potential value as a rule-out screening tool for MDR risk in critically ill AECOPD patients admitted to ICU.