Identification of risk factors in coronary bypass surgery

Abstract

In quality improvement in medical care one important aim is to prevent complications after a surgery and, particularly, keep the mortality rate as small as possible. Therefore it is of great importance to identify which factors increase the risk to die in the aftermath of a surgery. Based on data of 1,163 patients who underwent an isolated coronary bypass surgery in 2007 or 2008 at the Clinic of Cardiovascular Surgery in Düsseldorf, Germany, we select predictors that affect the in-hospital-mortality. A forward search using the wrapper approach in conjunction with simple linear and also more complex classification methods such as gradient boosting and support vector machines is performed. Since the classification problem is highly imbalanced with certainly unequal but unknown misclassification costs the area under ROC curve (AUC) is used as performance criterion for hyperparameter tuning as well as for variable selection. In order to get stable results and to obtain estimates of the AUC the variable selection is repeated 25 times on different subsamples of the data set. It turns out that simple linear classification methods (linear discriminant analysis and logistic regression) are suitable for this problem since the AUC cannot be considerably increased by more complex methods. We identify the three most important predictors as the severity of cardiac insufficiency, the patient's age as well as pulmonary hypertension. A comparison with full models trained on the same 25 subsamples shows that the classification performance in terms of AUC is not affected or only slightly decreased by variable selection. © Springer International Publishing Switzerland 2013.