Abstract
<title>Abstract</title> <p> Background Multidrug-resistant organisms (MDROs) such as methicillin-resistant <italic>Staphylococcus aureus</italic> (MRSA), vancomycin-resistant <italic>Enterococcus</italic> (VRE), and carbapenem-resistant <italic>Pseudomonas aeruginosa</italic> (CRPA) are major drivers of adverse outcomes in intensive care unit (ICU) patients with sepsis. At ICU admission, clinicians often must choose empiric antibiotics before culture and susceptibility results are available; inaccurate early risk assessment can delay active therapy or lead to unnecessary broad-spectrum antibiotic exposure. Methods We conducted a retrospective multicenter study using two publicly available ICU databases (eICU v2.0 and MIMIC-IV v3.1). Adult ICU admissions (age ≥ 18 years) with ICU length of stay ≥ 24 hours were included, and admissions with known MDRO infection or positive MDRO culture before ICU admission were excluded. The outcome window was restricted to clinically ordered cultures obtained from ICU admission through ICU day 7. We extracted 168 candidate predictors from the first 24 hours of ICU data; post-index variables, including ICU or hospital length of stay, were excluded from model development. The merged cohort (N = 45,141) was randomly split into a development set (70%, n = 31,599) and an independent validation set (30%, n = 13,542). Pathogen-specific models were evaluated by AUROC, decision curve analysis (0–50% threshold probabilities), and SHapley Additive exPlanations (SHAP). Results Within the prespecified early ICU outcome window, the prevalence of MRSA, VRE, and CRPA infections was 1.5% (n = 683), 0.7% (n = 296), and 0.2% (n = 90), respectively. In the validation set, AUROCs were 0.82 for MRSA, 0.85 for VRE, and 0.82 for CRPA. Decision curve analysis demonstrated positive net benefit across clinically relevant thresholds for all three models. SHAP analyses highlighted pathogen-specific model signals: eosinophil count, PaO2, and platelet count for MRSA; fibrinogen, lymphocyte count, and PaCO2 for VRE; and potassium, LDL cholesterol, and lymphocyte count for CRPA. Conclusions Explainable machine-learning models based on routinely available early ICU variables may support early MDRO risk stratification in sepsis and provide transparent, patient-level explanations. External and prospective validation will be required before clinical implementation. </p>