Multiscale Feature Enhancement and Bidirectional Temporal Dependency Networks for Arrhythmia Classification.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Cardiac arrhythmias, including premature beats and atrial fibrillation, present significant clinical challenges for detection.
• Existing deep learning models struggle with accurate differentiation between these specific arrhythmia types.

Purpose of the Study:

• To develop an advanced arrhythmia classification model overcoming limitations of single deep learning architectures.
• To enhance the accuracy of detecting premature beats and atrial fibrillation using a novel integrated approach.

Main Methods:

• A four-layer convolutional residual module with skip connections for multiscale ECG feature extraction.
• Multi-head self-attention mechanism to capture global feature correlations.
• Bidirectional long-term temporal dependency network for sequence contextual understanding.
• Dropout-regularized fully connected layer for six-type arrhythmia classification.

Main Results:

• Achieved an overall accuracy of 98.55% and an F1-score of 0.9531 on a fused dataset.
• Demonstrated superior F1-scores for premature beats (0.9916) and atrial fibrillation (0.9888), surpassing recent literature.
• Exhibited robust classification performance and effective identification of target arrhythmias.

Conclusions:

• The proposed model significantly improves the classification accuracy for premature beats and atrial fibrillation.
• This integrated deep learning approach shows strong potential as a supportive tool for automated ECG diagnosis.
• The model's performance highlights advancements in AI-driven cardiac arrhythmia detection.