Machine learning-based predictive models for perioperative major adverse cardiovascular events in patients with stable coronary artery disease undergoing non-cardiac surgery

Background Machine learning (ML)-based predictive models for perioperative major adverse cardiovascular events (MACEs) in patients with stable coronary artery disease (SCAD) undergoing non-cardiac surgery (NCS) have not been reported before. Methods Clinical data from 9171 consecutive adult patients with SCAD, who underwent NCS at the First Affiliated Hospital, Zhejiang University School of Medicine between January 2013 and May 2023, were used to develop and validate the prediction models. MACEs were defined as all-cause death, resuscitated cardiac arrest, myocardial infarction, heart failure and stroke perioperatively. Compare various resampling and feature selection methods to deal with data imbalance. A traditional logistic regression (the Revised Cardiac Risk index, RCRI) and nine ML models (logistic regression, support vector machine, Gaussian Naive Bayes, random forest, GBDT, XGBoost, LightGBM, CatBoost and best stacking ensemble model) were compared by the area under the receiver operating characteristic curve (AUROC) and the area under the precision recall curve (AUPRC). The calibration was assessed using the calibration curve and the patients’ net benefit was measured by decision curve analysis (DCA). Models were tested via 5-fold cross-validation. Feature importance was interpreted using SHapley Additive explanation (SHAP). Results Among 9171 patients, 514 (5.6%) developed MACEs. The XGBoost performed best in terms of AUROC (0.898) and AUPRC (0.479),which were better than the RCRI of AUROC (0.716) and AUPRC (0.185), Delong test and Permutation test P<0.001, respectively. The calibration curve of XGBoost performance accurately predicted the risk of MACEs (brier score 0.040), the DCA results showed that the XGBoost had a high net benefit for predicting MACEs. The top-ranked stacking ensemble model consisting of CatBoost, GBDT, GNB, and LR proved to be the best, with an AUROC value of 0.894 (95% CI 0.860-0.928) and an AUPRC value of 0.485 (95% CI 0.383-0.587). Using the mean absolute SHAP values, we identified the top 20 important features. Conclusion The first ML-based perioperative MACEs prediction models for patients with SCAD were successfully developed and validated. High-risk patients for MACEs can be effectively identified and targeted interventions can be made to reduce the incidence of MACEs. Lay Summary We performed a retrospective machine learning classification study of MACEs in patients with SCAD undergoing non-cardiac surgery to develop and validate an optimal prediction model. In this study, we analyzed the data missing mechanism and identified the best missing data interpolation method, while applying appropriate resampling techniques and feature selection methods for data imbalance characteristics, and ultimately identified 24 preoperative features for building a machine learning predictive model. Eight independent machine learning prediction models and stacking ensemble models were built, and the models were evaluated comprehensively using ROC curve, PRC curve, calibration plots and DCA curve. We have adopted a series of widely used machine learning algorithms and model evaluation techniques to build clinical prediction models, and achieved better performance and clinical practicability than the classical RCRI model, which has taken the first step to explore the research in this field. The prediction results based on the optimal machine learning model are interpretable, output the importance ranking and impact degree of the top 20 features of MACEs risk prediction, and are consistent with clinical interpretation, which is conducive to the application of the model in clinical practice.