To establish a decision tree model based on the synthetic minority over-sampling technique (SMOTE) algorithm for the prediction of post-operative 30-day death in patients undergoing surgery with cardiovascular events.

A total of 3 086 Chinese patients undergoing surgery with cardiovascular events (the history of ischemic heart disease and/or congestive heart failure) admitted to the Singapore General Hospital for operation from 2012 to 2016 were enrolled, and their clinical information, history of diseases and surgical scores were extracted. The original data was reconstructed by the SMOTE algorithm, and predictors were selected by best subset regression. Data was divided into a training group and a validation group by the ratio of 7 ∶ 3, of which the training group was used to establish the decision tree model and the validation group was used for internal verification.

The mortality rate was 3.0% (93/3 086) at 30 days after surgery and 4.5% (140/3 086) of patients were admitted to ICU at 24 h after surgery. The best subset regression analysis showed age > 75 years [odds ratio (OR) = 1.033, 95% confidence interval (CI) (1.024, 1.042), P < 0.001], anemia severity [OR = 1.368, 95%CI (1.211, 1.546), P < 0.001], chronic kidney disease stage > 2 [OR = 1.381, 95%CI (1.277, 1.494), P < 0.001], preoperative blood transfusion [OR = 4.496, 95%CI (3.268, 6.185), P < 0.001], surgical types [OR = 3.344, 95%CI (2.752, 4.064), P < 0.001], red blood cell distribution width > 15.7% [OR = 2.097, 95%CI (1.658, 2.652), P < 0.001] and American Society of Anesthesiologists classification > 2 [OR = 3.362, 95%CI (2.734, 4.135), P < 0.001] were risk factors for 30-day death after surgery in patients with cardiovascular events. The above seven predictors were selected to build a decision tree model. The results showed that the area under the receiver operating characteristic curve of the decision tree model was 0.853 [95%CI (0.837, 0.868), P < 0.001], and the sensitivity and specificity were 0.765 and 0.756 respectively in the training group; the area under the receiver operating characteristic curve of the decision tree model was 0.858 [95%CI (0.834, 0.882), P < 0.001], and the sensitivity and specificity were 0.938 and 0.612 respectively in the validation group, with good overall discrimination.

The risk of post-operative 30-day death in patients with cardiovascular events is an issue of unbalanced data classification, with minor cases for outcomes. In this study, the SMOTE algorithm was adopted avoiding the poor clinical applicability caused by the overfitting in conventional modeling. At the same time, the decision tree model presents visual, convenient and personalized characteristics, which is a useful clinical prediction tool for physicians.