Development and internal-external validation of a machine learning-based risk prediction model for multidrug...

This study aimed to develop and conduct internal-external validation of a machine learning-based risk prediction model for multidrug-resistant (MDR) infection in mechanically ventilated intensive care unit (ICU) patients with chronic obstructive pulmonary disease (COPD), so as to provide evidence for early identification of high-risk individuals, optimization of antimicrobial strategies, and reduction of infection-related adverse outcomes.

We retrospectively analyzed 477 mechanically ventilated COPD ICU patients from the MIMIC-IV database, which were divided into a training set (333 cases) and an internal validation set (144 cases) at a ratio of 7:3, with MDR infection as the endpoint. Feature selection was performed using the Boruta algorithm and least absolute shrinkage and selection operator (LASSO). Seven machine learning algorithms were adopted for model training and construction with 10-fold cross-validation. Model performance was evaluated by AUC, accuracy, sensitivity, specificity, F1 score, calibration curve and decision curve analysis. External validation was conducted using 371 independent single-center clinical cases. The optimal model was interpreted by Shapley additive explanations (SHAP) analysis, and a predictive nomogram was established.

Five core predictors were identified, including red blood cell count, hemoglobin, hematocrit, blood urea nitrogen and creatinine. The K-nearest neighbor (KNN) model outperformed other algorithms, with an internal validation AUC of 0.642 (95% CI: 0.518-0.766) and an external validation AUC of 0.639 (95% CI: 0.508-0.769), accompanied by favorable accuracy, calibration and clinical net benefit. SHAP analysis indicated that hematocrit was the primary predictive factor; elevated red blood cell-related indexes increased MDR risk, while elevated blood urea nitrogen and creatinine decreased the risk.

MDR infection in mechanically ventilated COPD ICU patients is closely associated with five routine laboratory indicators. The established KNN model and nomogram show favorable predictive value and generalizability, which can achieve rapid individualized risk quantification and serve as an auxiliary tool for early risk stratification and rational antimicrobial selection. It should be emphasized that this model is only used for preliminary screening and cannot replace microbial culture and drug susceptibility testing as the gold standard for definitive diagnosis.