The clinical applications of ensemble machine learning based on the Bagging strategy for in-hospital mortality of coronary artery bypass grafting surgery.

BACKGROUND

Machine learning (ML) has excelled after being introduced into the medical field. Ensemble ML models were able to integrate the advantages of several different ML models. This study compares the ensemble ML model's and EuroSCORE II's performance predicting in-hospital mortality in patients undergoing coronary artery bypass grafting surgery.

METHODS

The study included 4,764 patients from three heart centers between January 2007 and December 2021. Of these, 3812 patients were assigned to the modeling group, and 952 patients were assigned to the internal test group. Patients from other two heart center (1733 and 415 cases, respectively) constituted the external test group. The modeling set data are trained using each of the three ML strategies (XGBoost, CatBoost, and LightGBM), and the new model (XCL model) is constructed by integrating these three models through an ensemble ML strategy. Performance of different models in the three test groups comparative assessments were performed by calibration, discriminant, decision curve analysis, net reclassification index (NRI), integrated discriminant improvement (IDI), and Bland-Altman analysis.

RESULTS

In terms of discrimination, the XCL model performed the best with an impressive AUC value of 0.9145 in the internal validation group. The XCL model continued to perform best in both external test groups. The NRI and IDI suggested that the ML model showed positive improvements in all three test groups compared to EuroSCORE II.

CONCLUSIONS

ML models, particularly the XCL model, outperformed EuroSCORE II in predicting in-hospital mortality for CABG patients, with better discrimination, calibration, and clinical utility.