Toward Better Risk Stratification for Implantable Cardioverter-Defibrillator Recipients: Implications of Explainable Machine Learning Models

Figures

• 

  Figure 1. Study flow diagram. ICD, implantable cardioverter defibrillation.

• 

  Figure 2. Comparison of C-index between CPH and ML algorithms for all-cause death (A) and first appropriate shock (B) in the test set. For predicting all-cause death, the C-index of CPH, EN-Cox, RSF, SSVM, and XGBoost were 0.760 (95% CI 0.752–0.768), 0.771 (95% CI 0.763–0.779), 0.781 (95% CI 0.773–0.788), 0.767 (95% CI 0.759–0.775), and 0.794 (95% CI 0.786–0.802), respectively. For predicting shock, the C-index of CPH, EN-Cox, RSF, SSVM, and XGBoost were 0.611 (95% CI 0.604–0.618), 0.608 (95% CI 0.601–0.615), 0.589 (95% CI 0.581–0.597), 0.621 (95% CI 0.613–0.628), and 0.588 (95% CI 0.580–0.596), respectively. CI, confidence interval; CPH, Cox proportional hazards regression; other abbreviations as in Table 2.

• 

  Figure 3. Model interpretability using SHAP values. CPH model (A), XGBoost (B) for prediction of death; CPH model (C), SSVM (D) for prediction of appropriate shock. The X-axis stands for SHAP value, and the predictor lies orderly on the Y-axis according to their importance (the higher in position, the more important). Only the top 10 important predictors are left in the plot. Each point on the summary plot represents a single predictor of an individual. Overlapping points are jittered in the y-axis. Red and blue colors respectively indicate higher and lower values of a predictor. For example, a high value of NT-proBNP increases the risk score of death in the CPH model predicting death. In other words, it is a risk factor. Conversely, a high value of diastolic blood pressure reduces the death risk score, making it a protective factor. SHAP, SHapley Additive exPlanations; other abbreviations as in Figure 2 and Table 1.

• 

  Figure 4. SHAP force plots for XGBoost predicting all-cause death (A) and SSVM predicting shock (B) in a single patient. Red and blue bars respectively represent the positive and negative effects of each predictor contributing to the occurrence of the outcome. The extent of impact is represented by the size of the bar. Bold values represent the predicted risk scores. Of note, values have been transformed and raw values can be seen in Table S2. Abbreviations as in Table 1, Table 2, and Figure 3.

• 

  Figure 5. Cumulative incidence curves of all-cause death (A) and first appropriate shock (B) by risk stratum.

• 

  Figure 6. The construction of bi-dimensional risk profiles. (A), The cumulative incidence curves. (B), Bar plot of the annual incidence rate. The X-axis and Y-axis represent the three increasing risk strata of appropriate shock and all-cause death, respectively. A 3*3 risk profile was built accordingly. As a result, three different treatment strategies might be considered.