Generalizable Hybrid Wavelet-Deep Learning Architecture for Robust Arrhythmia Detection in Wearable ECG Monitoring.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Electrocardiogram (ECG) analysis is crucial for diagnosing heart conditions.
• Accurate and efficient ECG rhythm classification is challenging due to signal variability and noise.
• Deep learning offers promising avenues for automated ECG interpretation.

Purpose of the Study:

• To develop and evaluate a progressive deep learning framework for ECG rhythm classification.
• To investigate the effectiveness of combining time-frequency representations (scalograms) with hand-crafted features.
• To assess the performance and efficiency of various deep learning architectures for ECG analysis.

Main Methods:

• ECG signals were transformed into scalograms and processed by Vision Transformer (ViT) and other architectures.
• Scalograms were fused with scattering and statistical features for enhanced robustness.
• Principal Component Analysis (PCA) was applied for feature dimensionality reduction.
• Training-time augmentations were used to address class imbalance.

Main Results:

• Vision Transformer (ViT) achieved high accuracy (0.8590) using pure image-based ECG analysis.
• FusionViT with fused features yielded the best performance (accuracy = 0.8623, F1-score = 0.8528).
• Fusion ResNet-18 provided a balance between accuracy and inference efficiency (0.016 s per sample).
• PCA reduced feature dimensionality while maintaining competitive performance.

Conclusions:

• The proposed framework demonstrates high accuracy, robustness, and efficiency for practical ECG rhythm classification.
• The combination of visual and statistical features, with optional PCA reduction, is effective.
• The framework is suitable for real-time monitoring on edge devices like wearables and mobile health apps.