Machine learning models for 1-year mortality prediction in hypertensive patients undergoing coronary artery bypass grafting surgery

Abstract

Abstract Background A significant proportion of patients undergoing coronary artery bypass grafting (CABG) surgery has a variety of comorbidities such as hypertension. Hypertensive patients tend to have poorer prognoses and higher mortality from acute coronary syndrome in short- and mid-term follow-ups. Machine learning (ML) prediction methods are now widely used in different clinical settings, operating more accurately and efficiently than clinicians and traditional scores in many instances. Purpose This study was designed to develop and evaluate the ML prediction models of 1-year mortality in hypertensive patients undergoing CABG in our institute during 2005–2015. Methods Several baseline and procedural characteristics of CABG hypertensive cases were collected. Train and test data were divided with the proportion of 70:30. Due to a high 1-year survival rate, the synthetic minority oversampling technique (SMOTE) was implemented. In addition, the optimal threshold for determination of sensitivity and specificity was found using 10-fold cross-validation in the train data. After performing feature selection using the random forest method, prediction models were developed using: Support Vector Machine (SVM), Naïve Bayes (NB), Logistic Regression (LR), K-Nearest Neighbors (K-NN), Extreme Gradient Boosting (XGB), and Random Forest (RF) algorithms. Area Under the ROC (Receiver Operating Characteristic) Curve (AUC) in addition to sensitivity, specificity, and accuracy were used to assess the performance of the models. Results Among 8943 hypertensive CABG patients with 1-year follow-up, the mean age was 68.27±9.29 years, of whom 303 died during the study period. Eleven features from a total of 26 were chosen to train the models. The most important prediction features being selected were total ventilation time, ejection fraction, triglyceride, and age. LR (AUC=0.819) and XGB (AUC=0.812) outperformed RF (AUC=0.804), NB (AUC=0.791), SVM (AUC=0.715) and KNN (AUC=0.718). The LR model demonstrated the highest specificity (83.00%) and accuracy (82.37%) while the XGB model was the most sensitive one (88.00%). Conclusion An applicable prediction system can help clinicians' decision-making for the risk of 1-year mortality in hypertensive CABG patients which may have higher death rate compared with normal population. Although LR, RF, and XGB performance were almost similar and more favourable than the other three, LR remained the most promising predictor, given its AUC. Funding Acknowledgement Type of funding sources: None.