Risk prediction of inappropriate implantable cardioverter-defibrillator therapy using machine learning

Figures

• 

  Figure 1. Feature importance of extra trees classifier on train-validation data. Feature importance on train-validation data in predicting inappropriate implantable cardioverter-defibrillator (ICD) therapy using the extra-trees classifier. The feature importance was calculated in decision tree models, which was obtained by summing up the Gini impurity reductions for that feature across all the nodes in which it is used for splitting. The features were sorted in order to value of the importance. * amp amplitude.

• 

  Figure 2. SHAP method on extra trees classifier. Interpretation of feature importance using the SHapley Additive exPlanation (SHAP) method for Extra trees classifier on the train-validation data, which is based on “the game optimal Shapley values”. Blue color points indicate patients who did not receive inappropriate ICD therapy, and those who did (red points). The Shapley value of each feature is displayed on the x-axis (defined as the SHAP value), where a large (right side) value indicates a positive contribution to the model. The six features showing the highest absolute SHAP value were a history of atrial arrhythmia, cardiac resynchronization therapy (CRT), ischemic cardiomyopathy (ICM), V5 R-wave amplitudes, V3 ST level at J point, and verapamil administration. The interpretation of SHAP method is briefly explained in Supplemental file S2. * amp amplitude.

• 

  Figure 3. Cardi35 score for predicting inappropriate ICD therapy. (a) Definition of Cardi35 score and cut-off value of receiver operating characteristic curve analysis. On train-validation data, the predictive scores, calculated using this linear combination formula, was referred to as the Cardi35 score, which includes CRT, atrial Arrhythmia, DM, ICM, and V3 STJ and V5R amp. The six features were selected from the search the best number of features: Recursive feature elimination with cross validation method (RFECV). The score was weighted coefficient (= “point” in the table) derived from evaluation of features by Extra trees classifier on train-validation data. The score takes the values from − 3.0 to + 7.5. The cut-off value was + 5.0, with sensitivity and specificity of 0.732 and 0.920, respectively. The AUROC was 0.826 and the 95% confidence interval (CI) was 0.745–0.906. Asterisk area under the receiver operating characteristic curve. (b) Validation of Cardi35 score on the test data. The prognostic value of Cardi35 was validated on the test data. Hazard ratio of this score was calculated as 1.62 by Cox regression analysis (95% confidence interval: 1.32–1.98, p < 0.001). Furthermore,…