Uncertainty-Aware Ensemble Learning for Localizing Arrhythmia Origins from ECG

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Radio-frequency (RF) ablation is a primary treatment for idiopathic ventricular arrhythmias (IVAs), necessitating precise identification of the arrhythmia's source.
• Current invasive cardiac mapping systems for localizing IVAs are time-consuming and increase procedural risks.
• Non-invasive localization using electrocardiogram (ECG) signals is desirable but challenging due to the complex, non-linear relationship with arrhythmia origin and signal noise.

Purpose of the Study:

• To develop an advanced framework for accurate, non-invasive localization of IVAs' origin from ECG signals.
• To address the challenges of noise and non-linearity in ECG data for arrhythmia mapping.
• To improve upon existing methods for identifying arrhythmia sources, reducing procedural time and risk.

Main Methods:

• Proposed an uncertainty-aware ensemble learning (UAEL) framework integrating base models (segmentation and localization networks) with a unified UAEL algorithm.
• Base models learn QRS complex representations reflecting intracardiac electrophysiology for initial localization.
• UAEL algorithm incorporates an uncertainty-aware loss function and model decision refinement to enhance accuracy and account for patient variability.

Main Results:

• The UAEL framework demonstrated superior performance in segmenting QRS complexes and localizing IVAs' origin compared to 11 baseline models.
• Evaluations on 462 patients across two datasets confirmed the framework's effectiveness.
• The proposed method showed strong robustness against ECG signal perturbations.

Conclusions:

• The UAEL framework offers a significant advancement in non-invasive localization of IVAs' origin from ECG.
• This approach enhances accuracy and robustness, potentially reducing procedural risks associated with RF ablation.
• The uncertainty-aware methodology effectively handles patient-specific variability and noisy ECG data.