Machine learning differentiates right posterior from right anterior accessory pathways using 6-lead electrocardiograms in dogs with ventricular preexcitation

Abstract

Objective: To develop a machine learning (ML) model to identify fiducial points on canine ECGs to localize right-sided accessory pathways as posterior or anterior during ventricular preexcitation (VPE).

Methods: ECG recordings with VPE and documented accessory pathway locations were preprocessed for a 1-dimensional U-net algorithm. A web-based platform (https://setpsi.com/AccessoryPathways/) was created. Training used approximately 70% of pooled beats from 16 of 27 dogs. Testing used approximately 30% of pooled beats from 11 of 27 dogs to assess accurate diagnosis of posterior versus anterior accessory pathways. Vectorcardiograms and mean electrical axis (MEA) were calculated to validate the ML model. Fiducial boundary correctness and beat identification accuracy were assessed with receiver operator characteristic curves and reported as area under the curve (AUC; 95% CI). A Mann-Whitney test was used to compare MEA methods (median; IQR).

Results: The ML algorithm was trained on 3,405 beats and tested on 1,984 beats. The model identified fiducial points P wave to delta wave (AUC, 0.957; 95% CI, 0.957 to 0.958) and delta wave/QRS (AUC, 0.965; 95% CI, 0.964 to 0.966), classified individual beats as posterior (AUC, 0.917; 95% CI, 0.915 to 0.920) and anterior (AUC, 0.948; 95% CI, 0.947 to 0.949), and determined pathway location in 82% (9 of 11) of test dogs. Vectorcardiograms of posterior pathways showed oval or elliptical loops with superior leftward vectors, while anterior pathways displayed complex figure-eight loops with inferior leftward vectors. The MEA differed (P < .01) between posterior (-23.7°; IQR, 39.5°) and anterior (61.1°; IQR, 9.8°) pathways. Both methods validated the ML model.

Conclusions: The ML model accurately localized accessory pathways in canine VPE.

Clinical relevance: ML will advance the ability to accurately diagnose VPE.

Keywords: OAVRT; accessory pathways; artificial intelligence; machine learning; ventricular preexcitation.