Application of machine learning to develop a predictive model for hyperbilirubinemia after cardiac surgery

Objective This study aimed to develop a machine learning-based model to predict hyperbilirubinemia after cardiac surgery. Methods A retrospective analysis was performed on 554 patients who underwent cardiopulmonary bypass surgery at the First Hospital of Sun Yat-sen University in 2021. The predictors of HB were determined by Least Absolute Shrinkage and Selection Operator (LASSO)regression, and eight different machine learning algorithms were constructed, including naive Bayes (NB), support vector machine (SVM), and decision tree (DT), random forest (RF), light gradient boosting machine (LightGBM), gradient boosting decision tree (GBDT), extra trees, and adaptive boosting (AdaBoost). Shapley additive explanation (SHAP) is used for interpretability analysis of the model. Results A total of 401 patients were enrolled, and 20 features were identified via LASSO regression. Among the 8 algorithms, the Area Under Curve (AUC)value of extra trees was 0.846, which was superior to those of the other models. The top 4 features of SHAP analysis were the preoperative total bilirubin level and international normalized ratio (INR). Other important risk factors included intraoperative red blood cell infusion and dexmedetomidine (DEX) use. Conclusions The extra-trees model is a good model for predicting the occurrence of hyperbilirubinemia after cardiac surgery. The most important risk factors for postoperative HB were an increase in total bilirubin and the INR before the operation, an increase in red blood cell transfusion during the operation and no use of DEX.