Atrial Fibrillation and Atrial Flutter Detection Using Deep Learning.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Atrial fibrillation (AFib) and atrial flutter (AFL) are common arrhythmias.
• Accurate detection from electrocardiogram (ECG) signals is crucial for patient management.
• Existing methods may lack efficiency or interpretability for real-world deployment.

Purpose of the Study:

• To develop a computationally efficient and robust deep learning model for AFib and AFL detection.
• To ensure the model's interpretability for clinical relevance.
• To evaluate the model's performance on diverse public ECG datasets.

Main Methods:

• A lightweight 1D ConvNeXtV2-based neural network was designed.
• The model was trained on multiple large-scale public ECG datasets.
• Performance was evaluated on independent test sets, including MIT-AFDB, MIT-ADB, and NST.
• Guided Grad-CAM was used for model interpretability analysis.

Main Results:

• The model achieved a state-of-the-art F1-score of 0.986 on the MIT-AFDB dataset.
• The network is computationally efficient, with 770 k parameters and 46 MFLOPs per 10s window.
• Visualizations confirmed attention to P-wave morphology and R-R interval irregularities.

Conclusions:

• The proposed lightweight neural network offers robust and interpretable detection of AFib and AFL from single-lead ECGs.
• Its efficiency and interpretability make it suitable for resource-constrained wearable and bedside monitoring devices.
• Future work will explore multi-lead ECGs and a wider range of arrhythmias.