Objective

2

To construct a bleeding risk prediction model for patients with coronary heart disease （CHD） using antithrombotic therapy based on bioelectrical impedance of twelve

jing

-well points， and to choose the best algorithm for the model.

Methods

2

A total of 1931 CHD patients accepted antithrombotic therapy were included， and they were divided into the exposure group where bleeding events were reported at the time of enrollment， and the control goup where there is no self-reported bleeding events. The electrical impedance of the twelve

jing

-well points were detected， which represented the electrical impedance of the twelve channels. The SHAP analysis framework was used to analyze the fit models with three machine learning algorithms， that is LR， KNN， and XGBoost. The model with the best performance among the three algorithms was chosen for further analysis. The weight and performance of each feature in the finally selected model were analyzed， and the average SHAP values of all sample points were calculated and sorted to determine the channels with the highest weight in the model， that were the most closely related ones to the bleeding events； at the same time， subgroup analysis was performed for different bleeding events based on the same model.

Results

2

Among the 1931 patients， there were 176 cases in the exposure group and 1755 cases in the control group. The top three bleeding events were gum bleeding （72 times）， subcutaneous bleeding （49 times）， and epistaxis （34 times）. There were statistically significant differences in the electrical impedance of the

jing

-well points of lung， large intestine， stomach， heart， small intestine， kidney， pericardium，

sanjiao

（三焦）， and gallbladder channels between the two groups （

P

＜0.05）. The XGBoost algorithm performed best among the three algorithms； the area under the curve （AUC） of the validation set was 0.728， and the AUC of the test set was 0.709， indicating good applicability of the external data set， and therefore XGBoost algorithm was selected for further analysis. The average SHAP values showed that the differences in the electrical impedance of the top five channels， which were small intestine channel （0.161）， large intestine channel （0.152）， lung channel （0.088）， pericardium channel （0.069）， and heart channel （0.068）， had significantly higher weight in the bleeding risk prediction model. Subgroup analysis regarding the gingival bleeding， subcutaneous bleeding， and epistaxis showed that the average SHAP value of large intestine channel （0.080） was the highest in gingival bleeding events， while the average SHAP value of lung channel （0.097） and that of small intestine channel （0.162） were the highest values in epistaxis events and subcutaneous bleeding events， respectively.

Conclusion

2

The bleeding risk prediction model for antithrombotic therapy in CHD patients can be constructed based on the bioelectrical impedance of the twelve

jing

-well points， that is， the twelve channels. The model performs best with the XGBoost algorithm， which indicates close relationships between epistaxis and lung channel， gingival bleeding and large intestine channel， subcutaneous bleeding and small intestine channel.