Classification of multi-lead ECG with deep residual convolutional neural networks

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning for electrocardiogram (ECG) interpretation is rapidly advancing.
• Accurate classification of multi-lead ECGs is crucial for clinical diagnosis.
• Existing methods face challenges with data from diverse sources and label inconsistencies.

Purpose of the Study:

• To develop and evaluate a novel deep residual neural network method for accurate multi-lead ECG classification.
• To address label inconsistencies and improve generalization across different ECG databases.
• To assess the performance of the proposed method in a competitive challenge setting.

Main Methods:

• Utilized ECG recordings from seven open databases (PhysioNet/Computing in Cardiology Challenge 2021).
• Pre-processed ECG data to ensure consistent sampling rates and corrected label inconsistencies.
• Developed five deep residual convolutional neural network models optimized with an asymmetric loss function.
• Implemented a label mask to filter potentially mislabeled data.

Main Results:

• Achieved competitive official challenge scores across various multi-lead ECG configurations (e.g., 0.54 for twelve-lead, 0.52 for six-lead).
• Ranked within the top 10 in the challenge for all tested lead configurations (e.g., 5th for twelve-lead, 3rd for six-lead).
• Demonstrated strong generalization capabilities on the hidden test set.

Conclusions:

• The proposed deep learning method effectively corrects differential labeling tendencies across databases.
• The models exhibit robust generalization for classifying multi-lead ECGs.
• The developed approach shows significant potential for real-world clinical applications in ECG interpretation.