Introduction To address the delayed identification of carbapenem-resistant Pseudomonas aeruginosa (CRPA), we developed an interpretable machine-learning (ML) model for early risk stratification. Utilizing routine complete blood count (CBC) and demographic data, this tool targets the critical 48–72 hour interval before final antimicrobial susceptibility results. Methods Data from 1,666 patients with P. aeruginosa infection (223 CRPA) at a primary center were retrospectively analyzed, alongside an independent external validation cohort (n=471). Following the least absolute shrinkage and selection operator (LASSO) regression on 32 variables, eight ML algorithms were trained. Model interpretability and clinical utility were evaluated using Shapley Additive Explanations (SHAP) and decision curve analysis (DCA). Eight ML algorithms were trained using 5-fold cross-validation and Bayesian hyperparameter optimization. To ensure reproducibility and handle class imbalance, fixed random seeds were set, and a sensitivity analysis using the Synthetic Minority Over-sampling Technique (SMOTE) was conducted. Model calibration was assessed using the Brier score. Results LASSO identified seven predictors: sex, age, mean corpuscular volume (MCV), hemoglobin (HGB), platelet-to-lymphocyte ratio (PLR), systemic inflammatory response index (SIRI), and intensive care unit (ICU) admission status. Among the evaluated algorithms, the random forest (RF) model achieved the best discrimination. The training area under the receiver operating characteristic curve (AUC) was 0.993; it achieved an average 5-fold cross-validation AUC of 0.929 ± 0.005. In the internal test set, it achieved an AUC of 0.837 (95% CI: 0.779–0.893), specificity of 0.972, and sensitivity of 0.507, with excellent calibration (Brier score = 0.084). The model retained strong performance externally (AUC: 0.898, specificity: 0.985, sensitivity: 0.600, Brier score: 0.073). SHAP analysis indicated that HGB was the most influential feature, inversely associated with CRPA risk. Decision curve analysis supported the clinical utility across threshold probabilities ranging from 15% to 65%. Discussion This clinlabomics-based RF model provides a rapid, low-cost adjunct for early CRPA stratification. Given its exceptionally high specificity (0.97) and modest sensitivity, it functions exclusively as a reliable clinical “rule-in” tool. Positive predictions can confidently guide early targeted therapy and strict infection control. However, negative predictions cannot safely rule out CRPA, emphasizing its role alongside standard empirical practices rather than as a standalone screening instrument.
Yu et al. (Thu,) studied this question.