Machine Learning Localization of Early Right Ventricular Activation Sites Using QRS Integral Features.

Abstract

<h4>Background</h4>Accurate non-invasive localization of right ventricular (RV) arrhythmia origins remains a challenge in electrophysiology. This study investigates the feasibility of using machine learning models based on 12-lead ECG QRS integrals to localize early RV activation sites.<h4>Methods</h4>A generic RV mesh was constructed from a CT scan, consisting of 277 triangular elements. Two cohorts were used: a development cohort with 8 patients and 227 known pacing sites, and a validation cohort with 3 patients and 34 pacing sites. Each pacing site was assigned to the centroid of a mesh element. QRS integrals (∫QRS) were computed from eight ECG leads and used as input features for support vector regression (SVR) models with radial basis function (RBF) and linear kernels. ∫QRS values were trimmed over varying durations (30-160 ms in 10 ms increments) to identify the optimal QRS integration window through bootstrapped cross-validation. Localization accuracy was assessed using Euclidean distance, RMSE, and R2 across both cohorts.<h4>Results</h4>The RBF SVR trained on the initial 60 ms QRS interval yielded the lowest mean localization error - 9.5 mm in the development set and 14.4 mm in the independent test set. In contrast, the linear SVR yielded a mean localization error of 16.6 mm on the development set and 17.0 mm on the independent test set, with more stable performance across QRS durations. Axis-specific analysis revealed superior predictive accuracy along the y- and z-axes, while the x-axis-corresponding to the septal-free wall direction-showed reduced performance. In a validation cohort (n = 34), the mean localization errors were similar between kernels, and the difference was not statistically significant (p = 0.31).<h4>Conclusion</h4>QRS-integral-based SVR models enable millimeter-scale localization of RV pacing sites from surface ECGs. While nonlinear models provide greater accuracy in anatomically complex regions, linear models offer robustness and simplicity. These findings support the clinical potential of ECG-driven machine learning for guiding RV arrhythmia localization and complementing traditional mapping approaches.