Hybrid machine learning to localize atrial flutter substrates using the surface 12-lead electrocardiogram.

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence

Background:

• Atrial flutter (AFlut) is a common re-entrant atrial tachycardia.
• Current invasive electrophysiological mapping and catheter ablation for AFlut often lack detailed mechanistic understanding, potentially increasing procedure duration.

Purpose of the Study:

• To develop and evaluate a machine learning algorithm for discriminating AFlut locations using non-invasive 12-lead electrocardiogram (ECG) signals.
• To classify AFlut into cavotricuspid isthmus-dependent (CTI), peri-mitral, and other left atrium (LA) types.

Main Methods:

• Utilized a hybrid dataset of 1769 ECG signals (in silico and clinical).
• Extracted 77 features and trained a decision tree classifier using a hold-out approach.
• Validated and tested the classifier on a clinical test set of 38 patients (114 ECGs).

Main Results:

• The classifier achieved 76.3% accuracy on the clinical test set.
• Sensitivity for CTI, peri-mitral, and other LA classes were 89.7%, 75.0%, and 64.1%, respectively.
• With majority voting across patient segments, CTI class classification reached 92% accuracy.

Conclusions:

• A machine learning classifier using only non-invasive ECG signals can potentially identify AFlut mechanisms and locations.
• This non-invasive method shows promise for aiding in the planning and personalization of AFlut treatments.