A SHAP-Based Interpretable Model for MDRO-Infection Risk in Chronically Multimorbid Middle-Aged and Elderly ICU Adults

Aims/Background: Multidrug-resistant organism (MDRO) infection poses a significant clinical challenge, resulting in poor outcomes for middle-aged and older intensive care unit (ICU) patients with chronic multimorbidity. While conventional prediction models can identify high-risk patients, they often lack transparency in clinical decision-making. To bridge this gap, we aimed to develop a prediction model that incorporates Shapley Additive Explanations (SHAP)–based explainability analysis. This approach aims to achieve accurate risk prediction while quantitatively assessing the contribution of each risk factor, thereby providing clear guidance for clinical intervention. Methods: We employed a hospital-based consecutive sampling strategy to enroll 314 middle-aged and older patients with chronic multimorbidity admitted to the intensive care units of three Grade A tertiary hospitals and one secondary hospital in Bengbu, Anhui, from May 2024 to May 2025. Based on standard diagnostic criteria for MDROs, patients were categorized into an MDRO infection group and a non-MDRO infection group. Twenty candidate variables potentially associated with MDRO infection were extracted from electronic medical record systems. Univariate analysis, Least Absolute Shrinkage and Selection Operator (LASSO) regression, and multivariable logistic regression were utilized to identify high-risk factors for MDRO infection. The model’s discriminative ability and calibration were evaluated using receiver operating characteristic (ROC) analysis and the Hosmer–Lemeshow goodness-of-fit test. Internal validation was performed through bootstrap resampling (1000 iterations). We compared logistic regression, Naive Bayes, support vector machine (SVM), decision tree, linear discriminant analysis (LDA), and quadratic discriminant analysis (QDA) models, selecting the optimal model for SHAP analysis. Results: (1) Univariate analysis revealed significant differences between the two groups in age, length of hospital stay, admission diagnosis, history of antibiotic use, blood transfusion history, catheter category and number, mechanical ventilation, tracheostomy, white blood cell count, C-reactive protein, and other indices. (2) Variables identified as significant in univariate analysis were further analyzed using LASSO regression, which pinpointed eight predictors with non-zero coefficients; these were included in multivariable logistic regression for refined selection. (3) ROC analysis and the Hosmer–Lemeshow test indicated strong model performance, with an area under the curve (AUC) of 0.906 (95% confidence interval [CI]: 0.896–0.912), sensitivity of 77.92%, and specificity of 88.46%. Bootstrap internal validation confirmed that the model retained satisfactory discriminative ability. (4) Among the models compared, logistic regression demonstrated high predictive accuracy along with superior interpretability and stability, making it the chosen model for detailed analysis. A nomogram was developed based on seven key predictors: age, length of hospital stay, admission diagnosis, history of antibiotic use, tracheostomy, C-reactive protein, and white blood cell count. (5) SHAP analysis revealed that history of antibiotic use, length of hospital stay, and age were the most significant factors for MDRO infection among ICU patients with chronic multimorbidity, with antibiotic exposure identified as the strongest contributor. Conclusion: The prediction model developed in this study demonstrates strong predictive performance for in-hospital MDRO infection in ICU patients with chronic multimorbidity. It serves as a preliminary tool for identifying high-risk patients and may facilitate early risk stratification and targeted prevention strategies.