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

Area of Science:

• Electrophysiology
• Medical Imaging
• Machine Learning

Background:

• Accurate non-invasive localization of right ventricular (RV) arrhythmia origins is a significant challenge in electrophysiology.
• This study explores the use of machine learning (ML) models with 12-lead ECG QRS integrals to identify early RV activation sites.

Purpose of the Study:

• To investigate the feasibility of using ML models based on QRS integrals for non-invasive localization of RV arrhythmia origins.
• To assess the accuracy of different support vector regression (SVR) models in pinpointing RV pacing sites.

Main Methods:

• A generic RV mesh was created from CT scans.
• QRS integrals (∫QRS) were computed from ECG leads and used as input for SVR models (RBF and linear kernels).
• Optimal QRS integration windows were identified using bootstrapped cross-validation, with localization accuracy assessed by Euclidean distance, RMSE, and R2.

Main Results:

• The RBF SVR model, using an initial 60 ms QRS interval, achieved the lowest mean localization error (9.5 mm in the development set, 14.4 mm in the test set).
• Linear SVR showed more stable performance across QRS durations but with higher mean errors.
• Validation cohorts showed similar localization errors between kernels, with no statistically significant differences.

Conclusions:

• QRS-integral-based SVR models allow for millimeter-scale localization of RV pacing sites using surface ECGs.
• Nonlinear models offer higher accuracy in complex regions, while linear models provide robustness.
• These findings highlight the clinical potential of ECG-driven ML for RV arrhythmia localization and complementing traditional mapping.