Comparison of multiple machine learning methods for predicting postoperative hyperglycemia in patients without diabetes undergoing cardiac surgery

Abstract

Background: Stress-induced hyperglycemia (SHG) represents a significant metabolic complication in non-diabetic cardiac surgery older adult patients, with substantial implications for postoperative outcomes. Despite its clinical importance, reliable predictive tools remain scarce. This study systematically compared the performance of logistic regression 5 s. advanced machine learning algorithms for SHG risk prediction in this vulnerable population. Patients and Methods: We conducted a retrospective cohort analysis of 600 patients (≥65 years) undergoing cardiac surgery at a tertiary medical center (January 2021–May 2025). Six clinically relevant perioperative variables were incorporated into five predictive models: logistic regression, Random Forest (RF), Gradient Boosting Machine (GBM), Adaptive Boosting (AdaBoost), and Extreme Gradient Boosting (XGBoost). Model performance was rigorously evaluated using AUC-ROC with 95% confidence intervals, sensitivity, specificity, positive (PPV) and negative predictive values (NPV), and precision. Results: The incidence of SHG in this cohort was 70.5%. Comparative analysis revealed logistic regression as the top-performing model (AUC 0.944, 95% CI 0.923–0.966), surpassing other algorithms: GBM (0.923, 0.902–0.952), 10GBoost (0.904, 0.890–0.941), AdaBoost (0.916, 0.871–0.936), and RF (0.877, 0.866–0.932). Moreover, the logistic model achieved optimal performance in sensitivity (94.5%), specificity (93.4%), PPV (97.7%), and NPV (96.8%). Conclusion: In contrast to more complex machine learning approaches, logistic regression demonstrated superior predictive accuracy for SHG in non-diabetic cardiac surgery older adult patients. Its exceptional performance metrics and clinical interpretability support its practical utility as an effective decision-support tool for perioperative risk stratification and management.