Evaluating antimicrobial resistance and clinical outcomes in surgical ICU using a machine learning perspective: a retrospective observational study

Abstract

Background: Antimicrobial resistance (AMR) poses a critical threat to patient outcomes in intensive care units (ICUs), complicating treatment regimens and elevating mortality. This study aimed to assess the prevalence and patterns of AMR, antibiotic utilization, and clinical outcomes among postoperative patients in a surgical ICU in southern Iran, and developed predictive models for clinically significant resistance (MDR/XDR). Methods: We conducted a retrospective study; 106 postoperative patients admitted to the surgical ICU between January 2022 and December 2023 were evaluated. Demographic and clinical data, antibiotic usage metrics (including Days of Therapy [DOT], Length of Therapy [LOT], and Antimicrobial-Free Days [AFD]), and microbial culture results were extracted from electronic health records. Resistance patterns were classified as minor, multidrug-resistant (MDR), extensively drug-resistant (XDR), or pan-drug-resistant (PDR). Predictive modeling was performed using an XGBoost classifier and a logistic regression (LR) baseline, with hyperparameter tuning, fivefold cross-validation, and SHAP (SHapley Additive exPlanations) analysis for feature importance. This exploratory, single-center study in a resource-limited setting highlights hypothesis-generating insights but is constrained by sample size and generalizability. Results: In this cohort of 106 postoperative surgical ICU patients (median age, 66 years; 63.2% male), hypertension (33.7%) and diabetes mellitus (26.9%) were the most common comorbidities. The median ICU stay was 14.5 days, with an all-cause in-hospital mortality rate of 91.5%. Extensive antibiotic exposure was observed, with median DOT and LOT of 29.5 and 14.5 days, respectively, and broad-spectrum antibiotics were administered in 96% of cases. Among 175 microbial entries, 145 (83.82%) were culture-positive, predominantly Gram-negative bacteria (71.72%), with E. coli (20%), Acinetobacter (17.24%), and Klebsiella (16.55%) as leading pathogens. Notably, 62.07% of isolates were MDR and 3.45% were XDR, while no pan-drug resistant strains were identified. The XGBoost model achieved a test ROC-AUC of 0.786 and mean cross-validation AUC of 0.896 ± 0.05, with 70% accuracy and a macro F1-score of 0.70. The LR baseline yielded a test AUC of 0.743 and 77% accuracy, showing higher sensitivity but lower specificity. SHAP analysis identified Gram-negative infection type, Gram-positive infection type, LOT, and age as the most influential predictors of resistance. Conclusion: Surgical ICU patients experienced high rates of MDR infections, prolonged antibiotic exposure, and elevated mortality. Machine learning, particularly XGBoost, showed promising potential in this exploratory context for early identification of high-risk patients, highlighting its role in guiding antimicrobial stewardship and empirical therapy in critical care settings, pending further validation.