Development and validation of machine learning models for glycemic variability in non-diabetic patients following cardiopulmonary bypass: a prospective observational study

There is a correlation between glucose variability (GV) after cardiopulmonary bypass (CPB) and major adverse events. Identifying early risk factors and developing a prediction model for preventing GV is crucial. No machine learning models have been developed for GV in non-diabetic patients during CPB cardiac operations. This study established six models: logistic regression (LR), random forest (RF), decision tree (DT), support vector machine (SVM), eXtreme gradient boosting (XGBoost), and categorical boosting (CatBoost). Each model was internally validated, and the SHAP method identified important variables. Among 360 non-diabetic patients, 213 (59.17%) developed GV in the ICU. The models showed AUC values from 0.7400 to 0.818 in the training set and from 0.6658 to 0.763 in the testing set. XGBoost performed best, with an AUC of 0.736, accuracy of 0.7798, sensitivity of 0.875, positive prediction value of 0.7778, F1-score of 0.8235, and Brier score of 0.2041. Postoperative insulin, BMI, intraoperative mean glucose, and CPB duration were crucial features. By combining XGBoost with SHAP, the developed models can be used to facilitate individualized risk evaluation, allowing timely intervention or targeted care.