Deep Learning Classification of Unipolar Electrograms in Human Atrial Fibrillation: Application in Focal Source...

Area of Science:

• Cardiology
• Medical Technology
• Artificial Intelligence

Background:

• Focal sources are key targets for atrial fibrillation (AF) catheter ablation.
• Identifying these sources using unipolar electrograms (EGM) is challenging due to complexity and volume.
• Automating focal source detection is crucial for improving AF ablation efficiency.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for automated detection of putative focal sources in AF.
• To assess the performance of the DL model against manual classification by cardiologists.

Main Methods:

• A residual convolutional neural network (DL) was trained on raw unipolar EGMs from 78 patients in the Focal Source and Trigger (FaST) trial.
• The DL model was trained to discriminate between FaST (focal source) and non-FaST sites.
• A gradient-based method was used for DL model interpretation.

Main Results:

• The DL model achieved a high area under the curve (AUC) of 0.904 (cross-validation) and 0.923 (testing).
• At 90% sensitivity, the DL model demonstrated 81.9% specificity and 82.5% accuracy in detecting FaST sites.
• The DL model's performance was comparable to cardiologists' manual re-classification (sensitivity 78%, specificity 89% vs. 78%, 79%).

Conclusions:

• A novel DL model can accurately and automatically classify focal source sites (FaST) using raw unipolar EGMs.
• The DL model's performance is on par with expert cardiologists, suggesting its utility in clinical settings.
• This automated DL approach has the potential to enhance the efficiency of real-time focal source detection for targeted AF ablation therapy.