ECG data dependency for atrial fibrillation detection based on residual networks

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Atrial fibrillation (AF) is a common arrhythmia linked to stroke and heart failure.
• Deep learning algorithms show promise for AF detection but suffer from data dependency.
• Performance degradation on external datasets is a significant challenge for current deep learning models.

Purpose of the Study:

• To investigate the data dependency of deep learning-based AF detection algorithms.
• To evaluate the impact of dataset source on model performance.
• To assess the effect of training data volume on mitigating data dependency.

Main Methods:

• Utilized three independent PhysioNet databases for training and testing.
• Employed Residual Neural Network (ResNet) models (ResNet 18, 34, 50, 152).
• Trained models on raw electrocardiogram (ECG) signals and evaluated on external datasets.

Main Results:

• High accuracy (98-99%) achieved when models were tested on their own training data.
• Significantly lower accuracy (53-92%) observed when tested on external datasets.
• Data dependency was confirmed, but increased training data volume reduced this effect.

Conclusions:

• Deep learning models for AF detection exhibit significant data dependency.
• Model performance is highly sensitive to the source of training and testing data.
• Larger training datasets can substantially improve model generalizability and reduce data dependency.