Objective::Older patients with comorbidity, such as coronary heart disease (CHD) and malignant gastrointestinal tumors, are at a high risk of bleeding events. However, risk prediction models based on risk factor assessment remain unclear. This study aimed to establish an individualized bleeding risk assessment system based on the analysis of 10-year inpatient clinical big data.Methods::Total clinical data of 56,819 patients with CHD and 25,988 patients with malignant digestive tract tumors (admitted from January 2008 to December 2017) were retrospectively collected at the First and Second Medical Centers of Chinese People’s Liberation Army General Hospital. Among them, 1307 patients with CHD and malignant digestive tract tumors were screened as the derivation cohort. The dependent variable was the occurrence of major clinical bleeding events. Baseline statistics and hypothesis tests of differences were performed for independent variables according to the occurrence of bleeding. Decision Tree, eXtreme Gradient Boosting (XGBoost), logistic regression, and random forest models were used for comparison. The accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC) were applied as the criteria for evaluating and verifying model performance. To evaluate this developed model, another cohort comprising 454 patients (admitted from January 2018 to December 2019) was prospectively enrolled as the validation cohort based on the same inclusion and exclusion criteria.Results::Among the 64 variables with <50% missing values, the recursive feature elimination method with a random forest model was used to screen the selected variables. The highest accuracy was obtained following the selection of 10 scalars, and the final model was constructed accordingly. XGBoost demonstrated the best performance comprehensively. The AUC-ROC of this model was 0.981, with an accuracy, sensitivity, and specificity of 0.939, 0.950, and 0.927, respectively. In the validation cohort, the AUC-ROC, accuracy, sensitivity, and specificity of the XGBoost model were 0.702, 0.718, 0.636, and 0.725, respectively. The rate of major bleeding events has a positive correlation with the bleeding risk score quintiles. To allow for convenient clinical application, a smartphone application was developed for easy access and calculation (http://fir.master-wx.com/sghr).Conclusion::We successfully established a risk model and score for predicting bleeding events in older patients with comorbidity, such as CHD and gastrointestinal cancer.

Objective::Older patients with comorbidity, such as coronary heart disease (CHD) and malignant gastrointestinal tumors, are at a high risk of bleeding events. However, risk prediction models based on risk factor assessment remain unclear. This study aimed to establish an individualized bleeding risk assessment system based on the analysis of 10-year inpatient clinical big data.Methods::Total clinical data of 56,819 patients with CHD and 25,988 patients with malignant digestive tract tumors (admitted from January 2008 to December 2017) were retrospectively collected at the First and Second Medical Centers of Chinese People’s Liberation Army General Hospital. Among them, 1307 patients with CHD and malignant digestive tract tumors were screened as the derivation cohort. The dependent variable was the occurrence of major clinical bleeding events. Baseline statistics and hypothesis tests of differences were performed for independent variables according to the occurrence of bleeding. Decision Tree, eXtreme Gradient Boosting (XGBoost), logistic regression, and random forest models were used for comparison. The accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC) were applied as the criteria for evaluating and verifying model performance. To evaluate this developed model, another cohort comprising 454 patients (admitted from January 2018 to December 2019) was prospectively enrolled as the validation cohort based on the same inclusion and exclusion criteria.Results::Among the 64 variables with <50% missing values, the recursive feature elimination method with a random forest model was used to screen the selected variables. The highest accuracy was obtained following the selection of 10 scalars, and the final model was constructed accordingly. XGBoost demonstrated the best performance comprehensively. The AUC-ROC of this model was 0.981, with an accuracy, sensitivity, and specificity of 0.939, 0.950, and 0.927, respectively. In the validation cohort, the AUC-ROC, accuracy, sensitivity, and specificity of the XGBoost model were 0.702, 0.718, 0.636, and 0.725, respectively. The rate of major bleeding events has a positive correlation with the bleeding risk score quintiles. To allow for convenient clinical application, a smartphone application was developed for easy access and calculation (http://fir.master-wx.com/sghr).Conclusion::We successfully established a risk model and score for predicting bleeding events in older patients with comorbidity, such as CHD and gastrointestinal cancer.