Development of a predictive model using automated machine learning for Carbapenem-resistant Organisms (CRO) infections in hospitalized patients.

Objective: This study aimed to develop a predictive model based on Automated Machine Learning (AutoML) to achieve early and precise warning of Carbapenem-resistant Organisms (CRO) infections in hospitalized patients, thereby providing technological support for proactive prevention and control. Methods: A retrospective cohort study design was employed. Data from 958 hospitalized patients with CRO and Carbapenem-susceptible Organisms (CSO) infections were collected at CR and WISCO General Hospital between January 2022 and June 2025. An AutoML framework based on an Improved Hannibal Barcid Optimization algorithm (IHBO) was proposed. The model was constructed through dual-stage optimization (feature selection and hyperparameter tuning) and compared against five traditional algorithms, including Logistic Regression (LR) and Support Vector Machine (SVM). Model performance was evaluated using metrics including the Area Under the Receiver Operating Characteristic Curve (AUC), the Area Under the Precision-Recall Curve (AUPRC), and the F1 score. Key feature contributions and interaction mechanisms were analyzed using Shapley Additive exPlanations (SHAP). Results: Regarding model performance, the IHBO-optimized AutoML model demonstrated superior discriminative ability and robustness in the independent test set compared to others. Its ROC-AUC reached 0.8941 and PR-AUC reached 0.0.8844, significantly higher than those of the other models. Simultaneously, it achieved an F1 score of 0.8114, with sensitivity and specificity of 0.7917 and 0.8392, respectively. Calibration analysis indicated this model had the highest accuracy in predicted probabilities (Brier Score = 0.134). Decision curve analysis confirmed its significant clinical net benefit across the 1-71% risk threshold range. Feature analysis identified five core predictors ranked by importance: Antifungal Medication Usage, ICU Length of Stay Stratification, APACHE II Score > 15, Aminoglycoside Antibiotic Exposure, and Indwelling Urinary Catheter Use. SHAP interaction analysis further revealed: (1) Antifungal use significantly increases CRO risk, especially in patients with APACHE II > 15; (2) ICU stay duration shows dose-response relationship with CRO risk, amplified when APACHE II > 15; and (3) Combined use of aminoglycoside and urinary catheter creates synergistic risk, indicating additive effect of multiple risk factors; (4) Triple high-risk combination (antifungal + ICU > 7 days + APACHE II > 15) shows highest SHAP values, requiring intensive infection control measures. Based on these findings, a clinical decision support tool integrating the key features was developed, enabling the visual output of individualized infection risk. Conclusion: The IHBO-AutoML model developed in this study overcomes the limitations of traditional static methods. It employs explainable machine learning to elucidate the core driver mechanism involving the synergy between antifungal use and critical illness, the biological feedback loop linking prolonged ICU stay and organ failure, and the spatially specific resistance evolution induced by the interaction of drugs and medical devices. The model provides a precise tool for proactive prevention and control, facilitating the transition in carbapenem-resistant organism management modes.