Background: 
Substrate mapping based on intracardiac electrograms (EGMs) is an established approach during ablation of scar-related ventricular tachycardia (VT). However, consistent definitions of abnormal EGM features do not exist resulting in operator-dependent bias of ablation targets. Therefore, we aimed to develop an algorithm based on machine learning for automated identification of VT substrate.

Methods: 
In patients with ischemic cardiomyopathy presenting for VT ablation, bipolar EGMs from high-density substrate 3-D maps acquired during sinus rhythm were retrospectively analyzed. Bipolar EGMs were labeled binary as part or not part of the critical VT isthmus. Time-frequency representation images were created as corresponding heat maps of each bipolar EGM using wavelet transformation and were used as input data for the convolutional neural network (CNN) AlexNet (MatLab, version R2020) to eliminate typical feature selection of EGMs, and its corresponding loss of information.

Results: 
A total of 83,405 bipolar EGMs (5.9% part, 94.1% not part of the critical isthmus) from n=12 maps of n=14 VTs (mean cycle length 384±82 ms) in n=12 patients (75% male, 65±6.9 years, left ventricular ejection fraction 33±9.4%) were analyzed. For training of the CNN, All EGMs from the critical isthmus (n=3000) and 3 times the number of EGMs from noncritical (n=9000) areas were used leading to an accuracy of 96% for correct EGM classification (12,000 images from 10 patients). Testing (9,457 images from 2 patients) of the CNN resulted in an accuracy of 92%, a sensitivity of 75% and a specificity of 95% avoiding false positive labeling of abnormal EGMs of the critical isthmus zone.

Conclusion: 
The presented machine-based deep learning model is a promising tool for automated operator-independent identification of abnormal bipolar EGMs associated with the critical isthmus during VT ablation. Additional studies with more patients including long-term follow-up are needed for improvement of automated isthmus detection.