Machine Learning Models With Prognostic Implications for Predicting Gastrointestinal Bleeding After Coronary Artery Bypass Grafting and Guiding Personalized Medicine: Multicenter Cohort Study

Figures

• 

  Figure 1. Flowchart of the study. The study process is summarized as follows: data were collected from multiple centers, patient admission features were extracted, and the study cohort was defined. A derivation cohort was formed from 3 centers and a prospective cohort, which were used for feature screening and model construction. The optimal model was selected using multiple evaluation metrics and externally validated with the Drum Tower and MIMIC cohorts. Finally, the model was developed into a user-friendly web page to facilitate clinical use and guide treatment decisions. Abbreviations: COM: combined method, including features appearing more than three times in the first four feature selection methods; K-Best: k-best feature selection; LASSO: least absolute shrinkage and selection operator; LR: logistic regression; MI: mutual information; MIMIC IV: medical information mart for intensive care IV; MLP: multilayer perceptron; NB: naive Bayes; RF: random forest; RFE: recursive feature elimination; ROC curve: receiver operating characteristic curve; SHAP: Shapley additive explanations; SVM: support vector machine; XGB: XGBoost (extreme gradient boosting).

• 

  Figure 2. Performance of machine learning models. (A) Heatmaps illustrating the predictive performance (model score) of different combinations of feature selection methods (rows) and classification algorithms (columns). (B) Receiver operating characteristic curves of various models following LASSO feature selection. Abbreviations: COM: combined method; K-Best: k-best feature selection; LASSO: least absolute shrinkage and selection operator; LR: logistic regression; MI: mutual information; MLP: multilayer perceptron; NB: naive Bayes; PRECISE: “predicting bleeding complications in patients undergoing stent implantation and subsequent dual antiplatelet therapy” score; RF: random forest; RFE: recursive feature elimination; SVM: support vector machine; XGB: XGBoost (extreme gradient boosting).

• 

  Figure 3. Evaluation of the optimal model and its interpretability. (A) Calibration plots of XGBoost (Extreme Gradient Boosting) in the derivation cohort and 2 independent external validation cohorts. The Brier score is reported in the lower-right legend (a smaller value indicates better calibration). (B) Comparison of the area under the receiver operating characteristic curve (AUROC) for XGBoost and the PRECISE-DAPT (predicting bleeding complications in patients undergoing stent implantation and subsequent dual antiplatelet therapy) score in the Drum Tower and MIMIC (Medical Information Mart for Intensive Care) validation cohorts. (C) Comparison of clinical decision curves for XGBoost and the PRECISE-DAPT score in the combined external validation cohort. (D) Bar chart of the mean absolute Shapley Additive Explanations (SHAP) value for each predictor.(E) SHAP summary plot. The SHAP value along the x axis serves as a standardized measure of feature influence within the model. Each row of critical features illustrates the contributions of all patients to the outcome, with dots in different colors: red representing high-risk values and blue representing low-risk values.

• 

  Figure 4. Web-based calculator of the XGBoost (Extreme Gradient Boosting) model for risk stratification and personalized medicine. The web-based risk calculator predicts risk accurately online and in real time, providing a convenient tool for clinical doctors. It utilizes 15 features, including patient age, comorbidities, and initial examination results upon admission, to predict the risk of gastrointestinal bleeding in cardiac surgery patients. Additionally, it provides individualized interpretations of risk factors, outputs the patient's risk level, facilitates risk stratification, and guides targeted medication use.