Optical Mapping-Validated Machine Learning Improves Atrial Fibrillation Driver Detection by Multi-Electrode Mapping.

Area of Science:

• Cardiology
• Biomedical Engineering
• Computational Biology

Background:

• Atrial fibrillation (AF) is often sustained by localized reentrant drivers.
• Current clinical detection of AF drivers using multielectrode mapping (MEM) relies on subjective interpretation of activation maps.
• Automating AF driver detection could improve objectivity and treatment efficacy.

Purpose of the Study:

• To investigate the application of machine learning to electrogram frequency spectra for automated detection of AF drivers using MEM.
• To compare the accuracy of machine learning-based driver detection with near-infrared optical mapping (NIOM) as a gold standard.

Main Methods:

• Simultaneous mapping of AF drivers in human atria using NIOM and 64-electrode MEM.
• Analysis of unipolar MEM and NIOM recordings using Fourier transform to generate frequency spectra.
• Extraction of 35 features from each spectrum for machine learning classification.

Main Results:

• Machine learning classification of AF driver electrograms improved with averaged features from neighboring electrodes.
• Inclusion of driver periphery electrodes enhanced classification metrics.
• The algorithm achieved an f1-score of 0.81 for higher-density MEM data, significantly outperforming lower-density MEM (0.66).
• The trained algorithm correctly identified 86% of driver regions with higher-density MEM arrays.

Conclusions:

• A machine learning model utilizing Fourier spectrum features can efficiently classify electrograms as AF driver or nondriver compared to NIOM.
• This NIOM-validated machine learning approach holds potential for improving the accuracy of AF driver detection in clinical settings.
• Enhanced accuracy in AF driver detection can lead to more effective targeted ablation treatments for patients.