Deep Learning Approach for Highly Specific Atrial Fibrillation and Flutter Detection based on RR Intervals

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Atrial fibrillation (AF) and atrial flutter (AFL) are common atrial arrhythmias associated with a heightened risk of embolic stroke.
• Current detection methods for AF are effective, but distinguishing between AF and AFL remains a challenge for cardiologists.
• Accurate differentiation between these arrhythmias is crucial for optimizing patient therapy and reducing recovery time.

Purpose of the Study:

• To develop and evaluate a deep neural network for classifying electrocardiogram (ECG) signals into sinus rhythm (SR), AF, or AFL.
• To assess the model's performance in distinguishing between AF and AFL using RR interval sequences.
• To explore the potential of deep learning in creating highly specific detection procedures for AF and AFL.

Main Methods:

• A deep neural network architecture combining convolutional and recurrent neural networks was designed.
• The model extracted features from sequences of RR intervals in long-term ECG signals.
• A 10-fold cross-validation strategy was employed for architecture selection and hyperparameter tuning.

Main Results:

• The model achieved 88.28% accuracy during cross-validation, with sensitivities of 93.83% (SR), 83.60% (AF), and 83.83% (AFL).
• On a separate test set, the model demonstrated 89.67% accuracy, with sensitivities of 97.20% (SR), 94.20% (AF), and 77.78% (AFL).
• The proposed deep learning model showed promising performance in classifying SR, AF, and AFL.

Conclusions:

• The developed deep neural network effectively classifies ECG signals into SR, AF, and AFL.
• The model's ability to distinguish between AF and AFL holds potential for more efficient therapeutic strategies.
• Further development in deep learning-based detection of AF and AFL is warranted to enhance specificity and clinical utility.