Meta-learning based Active Learning Approach for Computer-Assisted Pace-Mapping

Cardiac diseases are the leading cause of death worldwide, with ventricular tachycardia (VT) being a major contributor. Successful VT intervention requires precise localization of the abnormal activation source, typically achieved through pace-mapping ie pacing different sites in the heart and comparing the resulting ECGs to the clinical VT ECG. Pace-mapping is invasive and timeconsuming, motivating the search for more efficient alternatives. Active learning methods using Gaussian process (GP) surrogate models have shown promise in reducing data requirements, but they rely on initial labeled data and cannot transfer knowledge between tasks. We propose a meta-learning–based neural active surrogate approach that transfers the knowledge learned from one pacing localization task to subsequent tasks without requiring any seed labeled data. Evaluated across two conditions (sinus and infarcted) in a heart geometry, our method achieved an 94% and 70% reduction in localization steps compared to random search and GP-based Bayesian optimization, respectively.

In Computing in Cardiology