Uncertainty-Aware Ensemble Learning for Localizing Arrhythmia Origins from ECG.

Abstract

Radio-frequency (RF) ablation is the most widely adopted treatment for idiopathic ventricular arrhythmias (IVAs), requiring accurate localization of the IVAs' origin. However, localizing the IVAs' origin with the invasive cardiac arrhythmia mapping system is time-consuming, which increases the associated procedural risk. Thus, localizing the IVAs' origin from the non-invasive electrocardiogram (ECG) signal has emerged as an attractive method, despite the challenge of learning the non-linear relationship between noisy ECG signal and IVAs' origin. To address the challenge, we propose the uncertainty-aware ensemble learning (UAEL) framework that integrates a set of base models with a unified UAEL algorithm. The base model, each comprising a segmentation network and a localization network, effectively learns the representation of QRS complexes that reflects intracardiac electrophysiology, thereby enabling accurate localization of the IVAs' origin. The UAEL algorithm further enhances the localization performance by incorporating an uncertainty-aware loss function to account for patient-specific variability, and an uncertainty aware model decision to refine predictions from individual base models. Evaluations on 462 patients across two datasets demonstrate that the UAEL framework outper-forms 11 baseline models in segmenting QRS complexes and localizing IVAs' origin, with strong robustness to ECG perturbations.