Application of machine learning algorithms to predict heat-sensitive angina (HSA) attacks: a multicentric observational cohort study

Figure 1. Flow diagram of the study population. SAP, stable angina pectoris; DFG, Danzhu Fuyuan Granule, a novel traditional Chinese medicine developed from the appropriate addition and reduction of Si-Miao-Yong-An decoction.

Figure 2. Predictor selection using Boruta. HDL, high-density lipoprotein cholesterol; HCY, homocysteine; TBIL, total bilirubin; NO, Nitric Oxide; TNF-α, tumor necrosis factor-alpha; SBP, systolic blood pressure; DBP, diastolic blood pressure; TG, triglyceride; CRP, C-reactive protein; LDL, low-density lipoprotein cholesterol; BMI, body mass index; ApoA-I, Apolipoprotein A-I; ApoB, Apolipoprotein B; FBG, fasting blood glucose; TC, total cholesterol; UA, uric acid; MPA, moderate physical activity.

Figure 3. Performance comparison of seven models in the internal validation cohort. (A) Receiver operating characteristic (ROC) curves for each model in the internal validation cohort; (B) Lollipop Chart for area under the ROC curve (AUC) values with 95% confidence intervals for each model in the internal validation cohort. DT, Decision tree model; KNN, K-Nearest Neighbor model; LGBM, Light Gradient Boosting Machine model; LR, Logistic regression model; NB, Naive Bayes model; RF, Random forest model; XGB, Extreme gradient boosting model.

Figure 4. Comparative performance analysis of seven models in the internal validation cohort. (A) Multi-metric performance summary of each model for HSA in the internal validation cohort; (B) Decision curve analysis (DCA) curves for each model in the internal validation cohort. HSA, heat-sensitive angina; DT, Decision tree model; KNN, K-Nearest Neighbor model; LGBM, Light Gradient Boosting Machine model; LR, Logistic regression model; NB, Naive Bayes model; RF, Random forest model; XGB, Extreme gradient boosting model.

Figure 5. Performance comparison of seven models in the external validation cohort. (A) Receiver operating characteristic (ROC) curves for each model in the external validation cohort; (B) Lollipop Chart for area under the ROC curve (AUC) values with 95% confidence intervals for each model in the external validation cohort. DT, Decision tree model; KNN, K-Nearest Neighbor model; LGBM, Light Gradient Boosting Machine model; LR, Logistic regression model; NBM, Naive Bayes model; RF, Random forest model; XGB, Extreme gradient boosting model.

Figure 6. Comparative performance analysis of seven models in the external validation cohort. (A) Multi-metric performance summary of each model for HSA in the internal validation cohort; (B) Calibration characteristics analysis for each model in the external validation cohort. HSA, heat-sensitive angina; DT, Decision tree model; KNN, K-Nearest Neighbor model; LGBM, Light Gradient Boosting Machine model; LR, Logistic regression model; NBM, Naive Bayes model; RF, Random forest model; XGB, Extreme gradient boosting model.

Figure 7. Interpretable machine learning analysis of risk factors. (A) All samples and features are illustrated, with each row representing a feature and x-axis representing the SHAP value. The purple dots represent higher feature values, while the yellow dots represent lower feature values; (B) Ranking of variable importance based on the average value; (C) SHAP predictions for HSA samples; (D) SHAP predictions for non-HSA samples. Yellow arrows indicate a higher risk of HSA, while purple arrows indicate a lower risk of HSA. The length of the arrows helps visualize the degree of impact of the prediction, whereby the longer the arrow, the more significant the effect. DBP, diastolic blood pressure; TG, triglyceride; CRP, C-reactive protein; LDL, low-density lipoprotein cholesterol; BMI, body mass index; ApoA-I, Apolipoprotein A-I; ApoB, Apolipoprotein B; FBG, fasting blood glucose; TC, total cholesterol; UA, uric acid; MPA, moderate physical activity.