Background: Machine learning models have potential to identify non-intuitive and previously unrecognized relationships between standardized clinical variables and the clinical manifestation of pathophysiological conditions. We used machine learning to examine the association of Society of Thoracic Surgeons (STS) Database variables with the presence of clinically significant ischemic mitral regurgitation (IMR) in patients undergoing coronary artery bypass grafting (CABG). Methods: STS Database variables (n=53) served as predictors of clinically significant IMR in machine learning modeling of 7,005 patients extracted from our institutional STS Database [1996-2018] who underwent CABG only (negative class, n=6,642) or CABG plus mitral valve intervention (positive class, n=363). Data were randomly partitioned into training (5,604 total patients, 281 positive, 5,323 negative) and test sets (1,401 total patients, 82 positive, 1,319 negative). The Synthetic Minority Oversampling Technique (SMOTE) was employed to produce a balanced training set. Results: Machine learning models, including random forests (RF), support vector machines (SVM), logistic regression (LR), and deep neural networks (DNN), were tested. Following training, final models predicted class labels for the patients in the test set. The models predicted class labels with promising accuracy (area under the receiver operating characteristic curve (AUC) values: RF, 0.70; SVM, 0.80; LR, 0.79; DNN, 0.80). Conclusions: STS Database variables have a predictive association with the presence of clinically significant IMR in patients undergoing surgical revascularization. These readily available variables may have potential as predictive variables in future translational machine learning modeling to assist in directing surgical care.
- Coronary artery disease
- Ischemic mitral regurgitation (IMR)
- Machine learning