Transformer-based temporal sequence learners for arrhythmia classification.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Electrocardiogram (ECG) is vital for diagnosing cardiac arrhythmia.
• Traditional and deep learning methods have limitations in classifying heart signals.
• ECG signals possess a time-sequential nature, suitable for advanced sequence modeling.

Purpose of the Study:

• To propose a transformer-based model for accurate ECG arrhythmia classification.
• To leverage the high parallelism of transformer architectures for heart signal analysis.
• To adapt a pre-trained natural language processing transformer (DistilBERT) for biomedical signal processing.

Main Methods:

• ECG signals were denoised, segmented around the R peak, and oversampled for dataset balancing.
• A DistilBERT transformer model was utilized, skipping input embedding and employing positional encoding.
• A classification head was added to the transformer encoder output for final probability generation.

Main Results:

• The proposed transformer model demonstrated excellent performance in classifying various arrhythmias.
• Achieved 99.92% accuracy on the MIT-BIH dataset.
• Obtained precision, sensitivity, and F1 scores of 0.99, with a ROC-AUC score of 0.999.

Conclusions:

• Transformer-based models, particularly DistilBERT, are highly effective for ECG arrhythmia classification.
• The proposed method offers a promising advancement over traditional and existing deep learning techniques.
• This approach holds significant potential for improving cardiac arrhythmia diagnosis and management.