Contrastive pretraining improves deep learning classification of endocardial electrograms in a preclinical model

Area of Science:

• Computational electrophysiology
• Machine learning in medicine
• Cardiac electrophysiology

Background:

• Persistent atrial fibrillation (AF) is driven by rotors and focal ectopies.
• Machine learning (ML) models can identify these drivers but are limited by small datasets.
• Developing accurate ML models for AF driver detection is crucial for understanding and treating the condition.

Purpose of the Study:

• To enhance the accuracy of ML algorithms for detecting drivers in persistent AF.
• To investigate the effectiveness of unsupervised pretraining on large unlabeled electrogram datasets.
• To improve the performance of driver detection models using transfer learning.

Main Methods:

• A SimCLR-based framework was used for unsupervised pretraining of a residual neural network.
• The network was pretrained on 113,000 unlabeled canine AF electrograms.
• The pretrained network was fine-tuned for driver detection and validated using various data augmentation techniques.

Main Results:

• Pretraining significantly improved driver detection accuracy from 62.5% to 80.8%.
• The pretrained model maintained high accuracy even with a 30% reduction in training data.
• Gradient-weighted Class Activation Mapping confirmed the model's attention aligned with known driver regions.

Conclusions:

• Contrastive pretraining enhances the accuracy and robustness of AF driver detection algorithms.
• Transfer learning shows promise for improving electrogram-based analyses in clinical electrophysiology.
• This method offers a pathway to more effective ML applications in diagnosing and managing AF.