Visualization deep learning model for automatic arrhythmias classification.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Heart disease is a leading global health threat.
• Electrocardiography (ECG) is crucial for diagnosing cardiovascular conditions.
• Automating arrhythmia classification is vital due to increasing ECG data and cardiologist shortages.

Purpose of the Study:

• To enhance the accuracy of detecting abnormal ECG patterns.
• To develop an automated system for classifying cardiac arrhythmias.
• To improve the interpretability of deep learning models in ECG analysis.

Main Methods:

• A hybrid 1D Resnet-GRU deep learning model was developed for 12-lead ECG analysis.
• The focal loss function was employed to address dataset imbalance.
• Grad-CAM++ was utilized for class-discriminative visualization to enhance model transparency.

Main Results:

• The 1D Resnet-GRU model achieved an F1-score of 0.821 in classifying 9 types of arrhythmias.
• Grad-CAM++ provided insights into the model's predictions, aligning with clinical arrhythmia diagnosis.
• The method demonstrated effective feature selection and integration for end-to-end classification.

Conclusions:

• The proposed hybrid deep learning model offers a promising approach for automated arrhythmia classification using 12-lead ECG.
• The integration of Grad-CAM++ provides an explainable AI framework for clinical decision support.
• This method facilitates accurate and transparent diagnosis of cardiovascular diseases.