Automated inter-patient arrhythmia classification with dual attention neural network.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Electrocardiogram (ECG) analysis is crucial for diagnosing heart rhythm disorders.
• Class imbalance in ECG data significantly hinders accurate arrhythmia classification.
• Existing methods struggle with performance for underrepresented arrhythmia categories.

Purpose of the Study:

• To develop a robust method for arrhythmia classification that addresses significant category imbalance in ECG data.
• To improve the diagnostic efficiency of clinical classification of heartbeats.
• To mitigate the adverse effects of imbalanced datasets on classification performance.

Main Methods:

• A novel dual attention hybrid network (DA-Net) was constructed, integrating modified convolutional networks with channel attention (MCC-Net) and a sequence-to-sequence network with global attention (Seq2Seq).
• MCC-Net extracts refined local features from ECG heartbeats.
• Seq2Seq fuses these features, leveraging local and global attention mechanisms for enhanced feature extraction and contextual understanding.

Main Results:

• The DA-Net achieved 99.98% accuracy for five-category arrhythmia classification on the MIT-BIH database without data augmentation.
• Exceptional performance was observed across all classes, including high sensitivity, specificity, and positive predictive values.
• The model maintained high accuracy (99.54% and 98.91%) even in extreme cases with significantly smaller training sample sizes.

Conclusions:

• The proposed DA-Net effectively alleviates the negative impact of class imbalance in arrhythmia classification.
• The model achieves excellent and robust performance across all categories without requiring data augmentation.
• This study presents a novel approach for handling class imbalance, demonstrating significant potential for conditions with limited data.