Explainable localization of premature ventricular contraction using deep learning-based semantic segmentation of...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Predicting premature ventricular contraction (PVC) origin from preoperative electrocardiograms (ECGs) is crucial for effective catheter ablation.
• Current methods require further refinement for precise origin localization.

Purpose of the Study:

• To develop an explainable deep learning method for localizing PVC origin using 12-lead ECGs.
• To provide diagnostic support for clinical applications in catheter ablation therapy.

Main Methods:

• A deep learning-based semantic segmentation model was trained on 265 12-lead ECG recordings from patients with frequent PVCs.
• The model classified ECG segments into background, sinus rhythm, left ventricular outflow tract PVC (PVC-L), and right ventricular outflow tract PVC (PVC-R).
• A rule-based algorithm further classified recordings into PVC-L, PVC-R, or a 'Neutral' category for physician assessment.

Main Results:

• The method achieved high performance metrics, including accuracy (0.932 on private, 0.916 on public datasets) and F1-score (0.945 on private, 0.943 on public datasets).
• Performance on a public dataset showed competitive results compared to previous studies.
• A notable portion of recordings fell into the 'Neutral' category, indicating a need for physician review.

Conclusions:

• The study demonstrated the feasibility of using explainable deep learning for localizing PVC origin from 12-lead ECGs.
• The developed method shows promise for enhancing diagnostic support in clinical settings.
• Further refinement may improve the classification of challenging cases requiring physician assessment.