Background: Ventricular tachycardia (VT) is a major cause of sudden cardiac deaths. To treat VT with ablation therapy, pace-mapping is used to locate the ablation target by matching 12-lead ECG stimulation sites to observed VT incidents. The current pace-mapping process relies on clinicians' interpretations of 12-lead ECGs in a laborious process. AI-assisted pace-mapping can allow for faster localization of VT targets.
Objective: Previous strategies using Gaussian processes (GP) in Bayesian optimization (BO) have reduced the number of pace-mapping sites required to reach localization. However, this requires re-learning the GP for each VT target, disregarding previously stimulated ECGs. This study presents a novel BO framework, combining neural ensembles (ENN) with continual-learning (CL) strategies to allow for knowledge transfer across tasks, reducing the number of pace-mapping sites needed for multiple VT targets within and across patients.
Method: We train an ENN to predict the ECG of a pace-mapping site and use its output to decide subsequent pace-mapping sites. Once a VT target is localized and a new task begins, an additional ENN is created and appended to a sequence of voters from each prior task. Additionally, we train a separate NN to weigh each voter according to their performance on the current task. Thus, the sequence of voters provides only the information that is most relevant to the current task.
Results: We evaluated our method on two heart geometries, each with four infarcted and one healthy condition, containing 3701 and 3357 pairs of pace-mapping sites and corresponding ECG. Results showed on average 87.7% reduction in pace-mapping sites compared to without active guidance, and 56.1% reduction compared to GP-based BO. Conclusion: We propose a continual-learning Bayesian optimization framework for site localization of target ECGs that enables knowledge transfer across tasks and improves the efficiency of pace-mapping.