Transfer Learning for Improved Classification of Drivers in Atrial Fibrillation.

Area of Science:

• Cardiology
• Computational Biology
• Machine Learning

Background:

• Persistent atrial fibrillation (AF) is driven by complex electrophysiological mechanisms.
• Identifying these "drivers" is crucial for effective AF treatment.
• Current machine learning (ML) approaches are limited by small driver datasets.

Purpose of the Study:

• To enhance ML classifier performance for detecting AF drivers.
• To investigate the utility of unsupervised pretraining on large unlabeled electrogram datasets.
• To improve the accuracy of driver detection algorithms.

Main Methods:

• Utilized a SimCLR-based framework for unsupervised pretraining.
• Employed a residual neural network architecture.
• Trained on a large dataset of 113K unlabeled 64-electrode electrogram measurements.

Main Results:

• Pretraining significantly improved weighted testing accuracy compared to a non-pretrained network (78.6±3.9% vs 71.9±3.3%).
• Demonstrated the effectiveness of transfer learning for electrogram data analysis.
• Established a foundation for developing more accurate driver detection algorithms.

Conclusions:

• Unsupervised learning pretraining is a viable strategy to boost ML performance in identifying AF drivers.
• Transfer learning shows promise for analyzing endocardial electrogram datasets.
• This approach paves the way for improved diagnostic tools in cardiology.