Ventricular geometry–regularized QRSd predicts cardiac resynchronization therapy response: machine learning from crosstalk between electrocardiography and echocardiography

Abstract

Up to one-third of patients selected by current guidelines do not respond to cardiac resynchronization therapy (CRT), the aim of this study was to find out novel analytical approaches to improve pre-implantation CRT response prediction. Among 31 pre-implantation features of clinical, laboratory, electrocardiography (ECG), and echocardiography variables in a consecutive cohort of patients receiving a first-time CRT device (CRT-pacemaker or CRT-defibrillator), we developed a machine learning (ML) model with three classification algorithms (support vector machines (SVM), K nearest neighbors, and random subspaces) with the best features combination to predict CRT response. Three categorical variables, left bundle branch block (LBBB), nonischemic cardiomyopathy, and female gender, were independently associated with CRT responses. Among continuous variables, including septal wall thickness, posterior wall thickness, and relative wall thickness (RWT), could regularize ECG QRS duration (QRSd) and significantly enhance the correlation between QRSd and CRT response. The 3 ML algorithms in a total of 38 features combinations constantly recognized that the features combined with QRSd/RWT outperformed the combinations without it. For each of three algorithms, the triplet feature combination of QRSd/RWT, LBBB, and nonischemic cardiomyopathy repeatedly increased the classification rate more than 8%. The best performance for CRT response prediction occurred with SVM model, which proposed actual QRSd/RWT values that favored CRT responses in patients both with and without LBBB. Lower QRSd/RWT values were required for CRT responses in patients with ischemic cardiomyopathy compared to those with non-ischemic cardiomyopathy. ML from ventricular remodeling characteristics–regularized QRSd improves CRT response prediction.

Appendix

Supplemental methods

For SVM, radial kernel was used. Mahalanobis distance was used in KNN. For all of the three methods, cross-validation was applied to choose appropriate values of the parameters.

Cross-validation, a well-known model/parameter selection technique, was performed to select appropriate parameters and estimate the test classification error. More specifically, we randomly split the whole data into ten subsets. For each experiment, nine of subsets were used as training data and the remaining one was used as test data. In other words, the features and the corresponding labels of training data were used to learn the model and then this model was applied to the features of test data to predict the response. The predicted labels were then compared with the ground truth to compute the test classification rate. The subset for testing is then permuted and ten experiments are carried out. The report classification rate is the average over ten experiments.

The advantage of cross-validation is twofold. First, the best model and parameter can be determined using the training data. Second, there was no overlap between the training and test data, and the ground-truth labels of the test data were not used for predicting the labels. Therefore, the test classification rate computed through the above procedure was a good measure about how well the model performs on further unseen data. Furthermore, the data are split at random and the subset for testing is permuted. It is well known that the variation in the averaged test classification rate is small.