Ten-Year Multicenter Retrospective Study Utilizing Machine Learning Algorithms to Identify Patients at High Risk of Venous Thromboembolism After Radical Gastrectomy

Abstract

Purpose: This study aims to construct a machine learning model that can recognize preoperative, intraoperative, and postoperative high-risk indicators and predict the onset of venous thromboembolism (VTE) in patients.

Patients and methods: A total of 1239 patients diagnosed with gastric cancer were enrolled in this retrospective study, among whom 107 patients developed VTE after surgery. We collected 42 characteristic variables of gastric cancer patients from the database of Wuxi People's Hospital and Wuxi Second People's Hospital between 2010 and 2020, including patients' demographic characteristics, chronic medical history, laboratory test characteristics, surgical information, and patients' postoperative conditions. Four machine learning algorithms, namely, extreme gradient boosting (XGBoost), random forest (RF), support vector machine (SVM), and k-nearest neighbor (KNN), were employed to develop predictive models. We also utilized Shapley additive explanation (SHAP) for model interpretation and evaluated the models using k-fold cross-validation, receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis (DCA), and external validation metrics.

Results: The XGBoost algorithm demonstrated superior performance compared to the other three prediction models. The area under the curve (AUC) value for XGBoost was 0.989 in the training set and 0.912 in the validation set, indicating high prediction accuracy. Furthermore, the AUC value of the external validation set was 0.85, signifying good extrapolation of the XGBoost prediction model. The results of SHAP analysis revealed that several factors, including higher body mass index (BMI), history of adjuvant radiotherapy and chemotherapy, T-stage of the tumor, lymph node metastasis, central venous catheter use, high intraoperative bleeding, and long operative time, were significantly associated with postoperative VTE.

Conclusion: The machine learning algorithm XGBoost derived from this study enables the development of a predictive model for postoperative VTE in patients after radical gastrectomy, thereby assisting clinicians in making informed clinical decisions.

Keywords: gastrectomy; gastric neoplasms; machine learning; prediction model; risk factors; venous thromboembolism.

Figure 1

Flow diagram of patients included in the study.

Figure 2

The variable influence factor ranking plots of the four models. (A) Variable importance ranking diagram of the XGBoost model. (B) Variable importance ranking diagram of the RF model. (C) Variable importance ranking diagram of the SVM model. (D) Variable importance ranking diagram of the KNN model.

Figure 3

Evaluation of the four models for predicting VTE. (A) ROC curves for the training set of the four models. (B) ROC curves for the validation set of the four models. (C) Calibration plots of the four models. The 45-degree dashed line in each plot represents the ideal correspondence between the predicted (x-axis) and observed (y-axis) probabilities of complications. The closer the distance between the two curves, the higher the predictive accuracy. (D) DCA curves of the four models. The point of intersection between the red curve and the “All” curve represents the baseline or starting point, while the point of intersection between the red curve and the “None” curve indicates the decision node where the corresponding patients may derive benefit.

Figure 4

Internal validation of the XGBoost model. (A) ROC curve of the XGBoost model for the training set. (B) ROC curve of the XGBoost model for the validation set. (C) ROC curve of the XGBoost model for the test set. (D) External validation of the XGBoost model.

Figure 5

SHAP summary plot. The risk factors are ranked on the y-axis according to their significance, which is determined by the mean of their absolute Shapley values. The higher the risk factor appears on the plot, the more crucial it is for the model.

Figure 6

SHAP force plot. The explanatory variables are ordered along the horizontal axis based on the absolute value of their impact, with blue representing features that negatively affect disease prediction, as indicated by a decrease in SHAP values, and red representing features that positively affect disease prediction, as indicated by an increase in SHAP values. (A) Predictive Analysis of Patient I. (B) Predictive Analysis of Patient II. (C) Predictive Analysis of Patient III.