Developing an explainable machine learning model using body composition to predict cardiovascular mortality in initial dialysis patients: a multicenter study

Figures

• 

  FIGURE 1. Flowchart of the study participants. SVM indicates support vector machine; GBM, Gradient Boosting Machine; XGBoost, extreme gradient boosting; Adaboost, Adaptive Boosting; CatBoost, Categorical Boosting.

• 

  FIGURE 2. LASSO regression analysis for predictor selection. As shown in (A), the plot illustrates the relationship between the binomial deviance and log(λ). The vertical dashed lines clearly mark the two λ values selected based on the minimum deviance criterion and the one-standard-error rule. This approach helps balance model accuracy while reducing the risk of overfitting. (B) presents the trajectories of variable coefficients across the log(λ) sequence, with each colored line representing a different predictor. As λ increases, the coefficients of weakly relevant or irrelevant predictors progressively shrink toward zero, thereby achieving feature selection and ultimately retaining only the most important variables in the model.

• 

  FIGURE 3. Comparison of ROC curves for eight machine learning models in the internal and external validation sets. SVM indicates support vector machine; GBM, Gradient Boosting Machine; XGBoost, extreme gradient boosting; Adaboost, Adaptive Boosting; CatBoost, Categorical Boosting. (A) Receiver operating characteristic curve in internal validation. (B) Receiver operating characteristic curve in external validation.

• 

  FIGURE 4. Calibration curves of eight machine learning models in the internal validation set. The horizontal axis represents the predicted probability of CVD mortality, and the vertical axis shows the observed event probability based on actual outcomes. The dashed line indicates the ideal reference line, where predicted probabilities perfectly match the observed outcomes. The closer each model’s calibration curve is to the dashed line, the better its calibration performance. SVM indicates support vector machine; GBM, Gradient Boosting Machine; XGBoost, extreme gradient boosting; Adaboost, Adaptive Boosting; CatBoost, Categorical Boosting; CVD, cardiovascular disease.

• 

  FIGURE 5. Decision curve analysis of eight machine learning models in the internal and external validation sets. The black horizontal line represents the net benefit when all patients are assumed to be non-CVD deaths, while the light gray curve represents the net benefit when all patients are assumed to be CVD deaths. The area between each model’s colored curve and the light gray curve reflects its clinical utility. A larger net benefit indicates greater clinical decision-making value of the model within the corresponding range of risk thresholds. SVM indicates support vector machine; GBM, Gradient Boosting Machine; XGBoost, extreme gradient boosting; Adaboost, Adaptive Boosting; CatBoost, Categorical Boosting; CVD, cardiovascular disease. (A) Decision curve analysis in internal validation. (B) Decision curve analysis in external validation.

• 

  FIGURE 6. Calibration curves of eight machine learning models in the external validation set. The horizontal axis represents the predicted probability of CVD mortality, while the vertical axis shows the observed event probability based on actual outcomes. The dashed line indicates the ideal reference line, where predicted probabilities perfectly correspond to observed results. The closer each model’s calibration curve is to the dashed line, the better its calibration performance. SVM indicates support vector machine; GBM, Gradient Boosting Machine; XGBoost, extreme gradient boosting; Adaboost, Adaptive Boosting; CatBoost, Categorical Boosting; CVD, cardiovascular disease.

• 

  FIGURE 7. ROC curves of the CatBoost model for 1-, 3-, and 5-year CVD mortality in the external validation set. (A) 1-year ROC curve. (B) 3-year ROC curve. (C) 5-year ROC curve.

• 

  FIGURE 8. Global SHAP value explanation of predictive features. (A) The bar plot shows the mean SHAP values of each predictor, ranked in descending order of their contribution to the model. This reflects the relative importance of each variable in the model’s decision-making process. (B) The SHAP scatter plot presents the SHAP values (x-axis) for each patient in the training set. Each point represents an individual patient, with colors indicating the magnitude of the corresponding feature value (yellow for high values and purple for low values). A positive SHAP value indicates a positive effect of the feature on the predicted outcome, whereas a negative SHAP value indicates a negative effect.

• 

  FIGURE 9. Local interpretability: SHAP force plot illustrating individual prediction explanations. (A–D) present four representative patients from the testing set. Yellow bars indicate features that increase the predicted risk, whereas purple bars indicate features that decrease the predicted risk. E[f(x)] is the baseline value and f(x) is the final prediction.

• 

  FIGURE 10. Local interpretability: SHAP waterfall plot illustrating individual prediction explanations. (A–D) correspond to four different patients and show how variables move the prediction from the baseline E[f(x)] to the final output f(x). Rightward shifts indicate positive contributions, and leftward shifts indicate negative contributions.

• 

  FIGURE 11. Web-based visualization interface of the CatBoost model.