The Clinical Challenge
Postoperative patients in surgical ICUs represent a uniquely vulnerable population. They face heightened infection risks from surgical wounds, indwelling devices, and compromised immune function, often requiring broad-spectrum empirical antibiotics. When multidrug-resistant (MDR) organisms emerge, treatment options narrow dramatically, and mortality rises. In resource-constrained environments, clinicians must make critical decisions without the benefit of rapid molecular diagnostics, relying instead on clinical judgment and delayed culture results.
Our retrospective study at Namazi Hospital, a tertiary referral center, analyzed 106 postoperative surgical ICU patients admitted between January 2022 and December 2023. The cohort was critically ill: median age was 66 years, 63.2% were male, and the all-cause in-hospital mortality rate reached 91.5%. Patients experienced prolonged ICU stays (median 14.5 days) and extensive antibiotic exposure, with median Days of Therapy (DOT) of 29.5 days and Length of Therapy (LOT) of 14.5 days. Alarmingly, the median Antimicrobial-Free Days (AFD) was zero, indicating continuous antibiotic exposure throughout ICU admission.
Resistance Patterns: A Concerning Landscape
Analysis of 175 microbial entries revealed 145 culture-positive isolates, with Gram-negative bacteria predominating (71.72%). Escherichia coli (20%), Acinetobacter species (17.24%), and Klebsiella species (16.55%) emerged as the leading pathogens. Among these isolates, 62.07% exhibited multidrug resistance, and 3.45% were extensively drug-resistant (XDR). No pan-drug-resistant strains were identified, but the high MDR prevalence signals a critical need for targeted interventions.
Resistance rates varied significantly across antibiotic classes. Fluoroquinolones and third-generation cephalosporins demonstrated the highest resistance rates (68.27% and 71.72%, respectively), followed by folate pathway inhibitors (62.76%) and aminoglycosides (53.10%). Carbapenem resistance showed notable variation, with imipenem resistance at 22.76% compared to meropenem at 4.83%. These patterns underscore the importance of local antibiogram data in guiding empirical therapy decisions.
Machine Learning: A Promising Frontier
A key innovation of our study was the development of machine learning models to predict MDR/XDR infections using routinely available clinical data. We employed an XGBoost classifier with hyperparameter optimization via Optuna (100 trials), comparing its performance against a logistic regression baseline.
The XGBoost model achieved a test ROC-AUC of 0.786 and a mean fivefold cross-validation AUC of 0.896 ± 0.05, demonstrating robust discriminative ability. While the logistic regression model showed higher accuracy (77% vs. 70%), XGBoost exhibited superior balanced performance with a macro F1-score of 0.70. SHAP (SHapley Additive exPlanations) analysis identified Gram-negative infection type (mean absolute SHAP value: 1.76), Gram-positive infection type (1.36), length of therapy (1.31), and age (1.14) as the most influential predictors of resistance.
Critically, we implemented a deliberate feature selection strategy to prevent data leakage, excluding variables that directly indicate outcomes (such as specific microbial isolates) and focusing instead on patient demographics, comorbidities, surgical history, and antibiotic utilization metrics. This approach ensures clinical applicability by relying on data available at the point of care.
Clinical Implications and Future Directions
Our findings have several important implications for antimicrobial stewardship in critical care:
First, the prolonged antibiotic exposure observed—with median AFD of zero—reflects the challenges of managing critically ill postoperative patients in resource-limited settings. Implementing antibiotic timeout protocols and prospective audit-feedback mechanisms could help optimize prescribing practices.
Second, the high prevalence of MDR Gram-negative infections highlights the need for strengthened infection prevention and control measures, including enhanced hand hygiene, contact precautions, and environmental cleaning.
Third, machine learning models like XGBoost could serve as clinical decision support tools, helping clinicians identify high-risk patients who might benefit from escalated empirical therapy or alternative treatment strategies. This is particularly valuable in settings where rapid diagnostics are unavailable, and culture results require 48–72 hours.
A Call for Collaboration
This exploratory, single-center study provides hypothesis-generating insights but has inherent limitations, including sample size and generalizability. External validation across diverse patient populations and healthcare settings is essential before widespread implementation. I invite fellow researchers, clinicians, and data scientists interested in AMR surveillance, machine learning applications in medicine, and global health to connect and explore collaborative opportunities. Together, we can develop more robust, externally validated predictive models that support evidence-based antimicrobial stewardship worldwide.