Using machine learning methods for personalized assessment of the risk of death after coronary artery bypass surgery

Abstract

Relevance. The desire to improve and optimize the results of surgical treatment of coronary heart disease (CHD), along with the observed integration of artificial intelligence methods into healthcare, creates prerequisites for exploring the possibilities of machine learning meth Relevance. The desire to improve and optimize the results of surgical treatment of coronary heart disease (CHD), along with the observed integration of artificial intelligence methods into healthcare, creates prerequisites for exploring the possibilities of machine learning methods for predicting adverse outcomes after cardiac surgery. The purpose of our study was to evaluate and compare the accuracy of predicting death after CABG surgery using machine learning methods and the recommended cardiac risk assessment scale EuroSCORE 2. Materials and methods. Based on the analysis of depersonalized medical data on the outcomes of coronary artery bypass surgery in 2,826 patients with coronary artery disease (survivors — 2,785, deceased — 41), using machine learning methods (logistic regression (LR), LightGBM, XGBoost, CatBoost, boosting model), prognostic models were developed that assess the risk of intrahospital death after intervention. The forecasting efficiency of the obtained models was compared with the forecasting results of the EuroSCORE 2 scale. To evaluate the performance of the models, the metrics recommended for the analysis of unbalanced data were used: precision, recall, specificity, F1-measure, ROC-AUC. Results. The model developed with the help of LR had the maximum recall (0.88), but at the same time significantly overestimated the risk of death (precision — 0.03). F1-measure for the LR model was 0.06, ROC AUC — 0.77. Gradient boosting models (LightGBM, XGBoost, CatBoost), in comparison with LR, had higher indicators of precision, recall, specificity, F1-measures and AUC. At the same time, the best quality metrics were observed in the boosting model (BM), which combined LR and gradient boosting models. BM performance indicators: precision — 0.67, recall – 0,50, F1-measure — 0.57, specificity — 1.0, ROC-AUC — 0.85. The EuroSCORE 2 risk model showed extremely low efficiency in predicting death in the study sample: precision — 0.143, recall — 0.125, F1-measure — 0.133, specificity — 0.97, ROC-AUC — 0.47. Conclusion. Machine learning (ML) methods are promising in predictive analytics in cardiac surgery. In our study, predictive models based on ML showed an advantage in the accuracy of calculating the risk of hospital death after CABG in comparison with the classic EuroSCORE 2 model. To obtain an optimal risk model adapted to the conditions of application in the Russian Federation, largescale multicenter studies are needed.