Arrhythmia classification detection based on multiple electrocardiograms databases.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Cardiovascular diseases represent a leading global cause of mortality.
• Electrocardiogram (ECG) is crucial for non-invasive heart disease detection.
• Existing ECG databases suffer from limited sample sizes and imbalanced distributions, hindering effective neural network training.

Purpose of the Study:

• To address the scarcity and imbalance of ECG data for machine learning.
• To develop a unified and comprehensive ECG arrhythmia classification database.
• To evaluate the performance of neural networks trained on the proposed database.

Main Methods:

• In-depth analysis of three fine-labeled ECG databases.
• Extraction and unification of heartbeats with consistent sampling frequency.
• Development of a self-processing method for heartbeats.
• Formation of the Hercules-3 unified ECG arrhythmia classification database (80% training, 20% testing).

Main Results:

• A fully connected neural network trained on Hercules-3 achieved 98.67% accuracy for 16-class arrhythmia classification.
• The proposed data processing method enhanced classification recall by at least 6%.
• Significant improvements were observed in classification accuracy (≥4%) and F1-score (≥7%) compared to other methods.

Conclusions:

• The Hercules-3 database provides a robust solution for ECG arrhythmia classification challenges.
• The self-processing method for heartbeats is effective in improving classification metrics.
• This work facilitates the development of more practical and efficient neural network models for cardiac diagnostics.