Detection and classification of electrocardiography using hybrid deep learning models.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electrocardiography (ECG) is crucial for diagnosing cardiovascular diseases (CVDs).
• Accurate feature extraction from ECGs is essential for reliable automatic classification.
• Existing methods require enhancement for robust CVD detection.

Purpose of the Study:

• To develop and evaluate a hybrid Deep Learning (DL) model for enhanced ECG-based CVD classification.
• To improve the accuracy and robustness of automatic interpretation of ECG signals.
• To compare the performance of the proposed model against other DL techniques.

Main Methods:

• A hybrid DL approach integrating Convolutional Neural Network (CNN) and Variational Autoencoder (VAE) architectures was employed.
• The model was trained and validated on the extensive PTB-XL dataset, comprising 21,799 12-lead ECGs from 18,869 patients.
• Classification performance was evaluated across five super-classes and 23 sub-classes of CVD.

Main Results:

• The CNN-VAE model achieved a peak accuracy of 98.51%, with specificity at 98.12% and sensitivity at 97.9%.
• Excellent F1-score of 97.95% was recorded, alongside minimal false positive (2.07%) and false negative (1.87%) rates.
• Results were validated against annotations provided by expert cardiologists.

Conclusions:

• The proposed CNN-VAE model demonstrates superior performance in CVD classification compared to other DL methods.
• This hybrid architecture offers a promising tool for earlier CVD detection and treatment guidance.
• The model's ability to better characterize complex ECG signals facilitates improved clinical decision-making.