BOATMAP: Bayesian Optimization Active Targeting for Monomorphic Arrhythmia Pace-mapping.

Area of Science:

• Computational cardiology
• Machine learning in medicine
• Electrophysiology

Background:

• Ventricular tachycardia ablation requires precise localization of activation origins.
• Current machine learning approaches face challenges with patient variability and optimal data selection for training.
• There is a need for patient-specific models with interpretable guidance for pace-mapping.

Purpose of the Study:

• To introduce BOATMAP, a novel active learning approach for localizing ventricular activation origins from 12-lead ECGs.
• To provide clinicians with interpretable, real-time guidance for selecting optimal pacing sites.
• To develop a patient-specific model that accounts for individual cardiac anatomy and tissue properties.

Main Methods:

• Developed BOATMAP, an active learning strategy using Gaussian processes (GP) as a surrogate model.
• Inverted the input-output relationship to learn ECG similarity as a function of pacing site coordinates.
• Incorporated constraints to avoid pacing in non-viable regions like myocardial scars.

Main Results:

• Achieved an average localization accuracy of 3.9±3.6mm using only 8.0±4.0 pacing sites in simulations.
• Demonstrated accurate localization across various heart geometries and tissue properties.
• Provided real-time, interpretable guidance for clinical decision-making.

Conclusions:

• BOATMAP accurately and efficiently localizes ventricular activation origins using active learning.
• The approach enhances clinical decision-making by offering interpretable guidance for ablation therapy.
• BOATMAP represents a significant advancement in patient-specific modeling for electrophysiology.