Robust wave-feature adaptive heartbeat classification based on self-attention mechanism using a transformer model.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electrocardiography (ECG) is crucial for cardiovascular disease screening.
• Accurate heartbeat classification is vital for diagnosis and research.
• Existing methods require improvement for real-time applications.

Purpose of the Study:

• To develop an automatic heartbeat classification method using a transformer neural network with a self-attention mechanism.
• To enhance the accuracy and efficiency of ECG analysis for cardiovascular disease detection.

Main Methods:

• An adaptive heartbeat segmentation technique was employed to focus on time-dependent heartbeat representations.
• A 1D convolution layer embedded wave characteristics into symbolic representations.
• A transformer block with multi-head attention processed wave-embedding dependencies.
• The model was trained and fine-tuned using the MIT-BIH arrhythmia database (MIT-DB) and MIT-BIH supraventricular arrhythmia database (MIT-SVDB).

Main Results:

• The method achieved F1 scores of 0.86 for supraventricular ectopic beats and 0.96 for ventricular ectopic beats in the first group.
• An average F1 value of 99.83% was achieved in the second group, outperforming state-of-the-art methods.
• The transformer-based approach demonstrated superior performance in heartbeat classification.

Conclusions:

• A novel and effective transformer-based method for automatic heartbeat classification was proposed.
• This approach offers a promising solution for real-time ECG analysis and cardiovascular disease screening.
• The method is suitable for integration into wearable devices for continuous health monitoring.