Feasibility of Machine Learned Intracardiac Electrograms to Predict Postinfarction Ventricular Scar Topography

Area of Science:

• Cardiovascular Electrophysiology
• Computational Biology
• Medical Imaging

Background:

• Accurate scar delineation is crucial for guiding ventricular tachycardia catheter ablation.
• Scar patterns and distribution influence ablation success.
• Novel methods are needed to precisely identify scar tissue.

Purpose of the Study:

• To determine if scar patterns can be identified from intracardiac electrograms using computational signal processing.
• To investigate if a convolutional neural network can improve scar classification accuracy.
• To correlate electrogram data with histological scar patterns.

Main Methods:

• Sheep models with anteroseptal infarction were used.
• Histological models of postinfarction scar were created and coregistered with electroanatomic mapping.
• Intracardiac electrograms (bipolar and unipolar) were analyzed using signal processing features and a convolutional neural network (InceptionTime).

Main Results:

• Bipolar and unipolar voltage alone were poor scar classifiers.
• Signal processing features from bipolar electrograms achieved an AUC of 0.815 for no scar and lower for other scar types.
• A convolutional neural network trained on unipolar electrograms achieved high AUCs (up to 0.977) and accuracy (up to 0.959) for scar pattern classification.

Conclusions:

• Convolutional neural network analysis of unipolar electrograms offers excellent predictive value for scar pattern determination.
• Advanced computational analyses of electrogram data are essential for improving arrhythmogenic site identification.
• This approach holds promise for enhancing ventricular tachycardia ablation guidance.