A deep learning model predicted ablation targets with a mean AUC-ROC of 0.90 for LV versus RV differentiation and 0.88 for four-class location of PVCs.
Does a deep learning model using 12-lead ECG accurately predict ventricular ablation targets in patients undergoing catheter ablation?
A deep learning model using standard 12-lead ECG data can accurately predict ventricular ablation targets, potentially optimizing procedural planning and reducing intraoperative mapping time.
Absolute Event Rate: 0% vs 0%
Abstract Background Catheter ablation (CA) is an ultimate treatment option in the management of premature ventricular complexes (PVC)s and ventricular tachycardia (VT), both for symptom relief and to prevent progression into PVC-induced cardiomyopathy. Although CA is an effective treatment, the procedure is not without risk of complications. To optimize procedural planning and success and minimize complication risk, precise localization of PVCs is essential. Pre-procedural localization using 12-lead electrocardiogram (ECG) may reduce the need for extensive intraoperative mapping and subsequent intervention duration. Purpose The aim of this study was to develop a deep learning model, using only the 12-lead ECG, to predict ablation targets in patients undergoing CA. Methods In this retrospective cohort study, we included patients who underwent ventricular CA procedure at a tertiary care academic medical center between 2012 and 2024. A convolutional β-variational autoencoder (ß-VAE) was trained on over 200.000 PVCs across over 42.000 unique patients to learn the underlying factors that capture the variation of their morphology. Subsequently, these factors were used as input into an extreme gradient boosting (XGBoost) model to predict the ablation site either as a binary outcome—left ventricle (LV) or right ventricle (RV)—or as a four-class outcome: LV outflow tract (LVOT), LV non-LVOT, RV outflow tract (RVOT), or RV non-RVOT. Results In total, 415 procedures from 375 unique patients were used to train the predictive model. PVCs originated from the LVOT in 8.4% of cases, non-LVOT LV sites in 42.2%, the RVOT in 31.1%, and the non-RVOT RV sites in 8.9%. The XGBoost model achieved a mean AUC-ROC of 0.90 ± 0.03 on the validation set for distinguishing LV from RV targets; and a macro-AUC-ROC of 0.88 ±0.03 in the validation set for the four-class location. Shapley additive explanations (SHAP) analysis further confirmed the discriminatory power of specific latent features. Conclusion Our study demonstrates that a convolutional neural network, using a ß-VAE trained on 12-lead ECG data, can translate PVC beats into explainable latent factors. The subsequent prediction model achieved good performance to differentiate between left and right ventricular ablation targets.
Perez et al. (Thu,) reported a other. A deep learning model predicted ablation targets with a mean AUC-ROC of 0.90 for LV versus RV differentiation and 0.88 for four-class location of PVCs.