Prediction of response after cardiac resynchronization therapy with machine learning

Abstract

Aims: Nearly one third of patients receiving cardiac resynchronization therapy (CRT) suffer non-response. We intend to develop predictive models using machine learning (ML) approaches and easily attainable features before CRT implantation.

Methods and results: The baseline characteristics of 752 CRT recipients from two hospitals were retrospectively collected. Nine ML predictive models were established, including logistic regression (LR), elastic network (EN), lasso regression (Lasso), ridge regression (Ridge), neural network (NN), support vector machine (SVM), random forest (RF), XGBoost and k-nearest neighbour (k-NN). Sensitivity, specificity, precision, accuracy, F1, log-loss, area under the receiver operating characteristic (AU-ROC), and average precision (AP) of each model were evaluated. AU-ROC was compared between models and the latest guidelines. Six models had an AU-ROC value above 0.75. The LR, EN and Ridge models showed the highest overall predictive power compared with other models with AU-ROC at 0.77. The XGBoost model reached the highest sensitivity at 0.72, while the highest specificity was achieved by Ridge model at 0.92. All ML models achieved higher AU-ROCs that those derived from the latest guidelines (all P < 0.05). The effect size analysis identified left bundle branch block, left ventricular end-systolic diameter, and history of percutaneous coronary intervention as the most crucial predictors of CRT response. An online tool to facilitate the prediction of CRT response is freely available at http://www.crt-response.com/.

Conclusions: ML algorithms produced efficient predictive models for evaluation of CRT response with features before implantation. Tools developed accordingly could improve the selection of CRT candidates and reduce the incidence of non-response.

Keywords: Cardiac resynchronization therapy; Machine learning; Prediction; Predictive models; Response.