A hybrid hierarchical transformer model for ECG classification and age prediction

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Electrocardiograms (ECGs) are vital for cardiovascular diagnostics, necessitating advanced analytical models.
• ECG analysis involves hierarchical temporal scales, from waveform morphology to inter-beat intervals.
• Convolutional Neural Networks (CNNs) excel at local feature extraction, while Transformers capture long-range dependencies.

Purpose of the Study:

• To develop a hybrid hierarchical model, HiT-NeXt, that integrates CNNs and Transformers for comprehensive ECG analysis.
• To enhance the model's ability to capture both local morphological patterns and global temporal dependencies in ECG signals.
• To improve the accuracy of ECG abnormality classification and cardiological age prediction.

Main Methods:

• Developed HiT-NeXt, a hybrid model combining ConvNeXt-based CNNs for local feature extraction and hierarchical learning with Transformer blocks.
• Incorporated restricted attention windows within Transformer blocks to focus on relevant temporal contexts.
• Utilized relative contextual positional encoding to improve robustness to signal translations.

Main Results:

• HiT-NeXt demonstrated superior performance in ECG abnormality classification compared to state-of-the-art methods.
• The model achieved higher accuracy in cardiological age prediction than existing approaches.
• Performance of HiT-NeXt surpassed evaluations by human cardiologists in experimental tasks.

Conclusions:

• HiT-NeXt effectively integrates local and global feature extraction for advanced ECG analysis.
• The hybrid hierarchical approach offers a significant advancement in cardiovascular healthcare diagnostics.
• HiT-NeXt represents a promising tool for automated ECG interpretation and risk stratification.