Machine Learning-Based Prediction of Prolonged Mechanical Ventilation After Stanford Type A Aortic Dissection Surgery.

Purpose: Stanford type A aortic dissection (TAAD) is a life-threatening cardiovascular emergency associated with high morbidity and mortality. Prolonged mechanical ventilation (PMV) remains a common postoperative complication and is associated with delayed recovery and increased resource use. Although several logistic regression-based prediction models for PMV after TAAD surgery have been reported, evidence comparing conventional regression with machine learning approaches remains limited. This study aimed to develop and internally validate three prediction models, namely logistic regression, support vector machine (SVM), and extreme gradient boosting (XGBoost), for assessing the risk of PMV after TAAD surgery. Patients and Methods: A total of 511 patients with TAAD from Renmin Hospital of Wuhan University and Tongji Hospital affiliated with Tongji Medical College of Huazhong University of Science and Technology were retrospectively enrolled and randomly divided into a training set (75%) and a test set (25%). PMV was defined as invasive mechanical ventilation >48h from the end of surgery to the first successful extubation. Risk factors for PMV were screened by univariate analysis and least absolute shrinkage and selection operator (LASSO) regression. Predictive models were then constructed using logistic regression, SVM, and XGBoost. Model performance was evaluated using AUC, sensitivity, specificity, accuracy, recall, and F1-score, and assessed by calibration curves and decision curve analysis (DCA). SHAP analysis was performed for the final model. Results: PMV occurred in 159 patients, with an incidence of 31.1%. LASSO regression identified six predictors: preoperative lactate, cardiopulmonary bypass time, peak postoperative creatinine, postoperative plasma transfusion, lactate at the end of surgery, and pulmonary complications. In the test set, the AUCs of the SVM, logistic regression, and XGBoost models were 0.820, 0.805, and 0.784, respectively. Although SVM achieved the highest numerical AUC and showed relatively favorable calibration and clinical net benefit, the differences in AUC among the three models were not statistically significant. Therefore, SVM was selected for SHAP interpretation based on its balanced overall performance rather than statistical superiority in discrimination. Conclusion: The three models demonstrated acceptable internal performance for predicting PMV after TAAD surgery, with SVM showing a numerically higher AUC and favorable overall performance in discrimination, calibration, and decision curve analysis. The identified predictors are clinically interpretable, and the current value of these models may mainly lie in perioperative risk stratification and postoperative dynamic assessment, thereby providing potential support for postoperative respiratory management. However, because postoperative variables were included and external validation was not performed, the findings should be interpreted as preliminary and require prospective multicenter validation before routine clinical use.