AI-Assisted Pace-Mapping using Continual-Learning Methods in Bayesian Optimization

Ventricular tachycardia (VT) is a leading cause of sudden cardiac death, and effective treatment often relies on catheter ablation guided by pace-mapping. Traditional pace-mapping is a labor-intensive process requiring expert interpretation of ECGs. Recent methods using Gaussian process (GP)-based Bayesian optimization (BO) have improved efficiency by reducing the number of stimulation sites needed for localization, but they fail to transfer knowledge across different VT targets, requiring retraining for each new case. This study introduces a novel BO framework that integrates ensemble neural networks (ENN) with continual learning (CL) strategies to enable knowledge transfer across tasks. Allowing increased efficiency without the need for initial data and memory across tasks. Evaluated on one healthy and one infarcted setting of a heart geometry, our proposed method demonstrated an average 90% reduction in required pace-mapping sites compared to unguided approaches and a 65% reduction relative to GP-based BO.

In Computing in Cardiology