Electrocardiogram classification using TSST-based spectrogram and ConViT.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Electrocardiogram (ECG) is crucial for diagnosing arrhythmias and related cardiovascular diseases.
• ECG arrhythmia classification presents significant challenges.
• Existing methods often struggle with data imbalance and feature extraction.

Purpose of the Study:

• To develop a novel deep learning model for accurate ECG arrhythmia classification.
• To integrate the strengths of Convolutional Neural Networks (CNN) and Vision Transformers (ViT).
• To address challenges like class imbalance and enhance feature extraction for improved diagnostic accuracy.

Main Methods:

• A novel deep learning model, Convolutional Vision Transformer (ConViT), combining CNN and ViT architectures.
• Gated Positional Self-Attention (GPSA) layers to integrate convolutional inductive bias with attention mechanisms.
• Time-Reassigned Synchrosqueezing Transform (TSST) for enhanced time-frequency feature extraction.
• SMOTE algorithm for data augmentation and Focal Loss (FL) for minority-class weighting to handle data imbalance.

Main Results:

• The proposed ConViT model achieved an overall accuracy of 99.5% on the MIT-BIH arrhythmia database.
• High performance metrics, including specificity, F1-Score, and Matthews Correlation Coefficient (MCC) exceeding 94% for supraventricular ectopic beats (S) and ventricular ectopic beats (V).
• Demonstrated superiority over existing methods in ECG arrhythmia classification.

Conclusions:

• The ConViT model offers a powerful and accurate solution for ECG arrhythmia classification.
• The integration of advanced deep learning techniques and signal processing methods significantly improves diagnostic capabilities.
• This approach holds promise for enhancing the early detection and management of cardiovascular diseases.