Generalising electrocardiogram detection and delineation: training convolutional neural networks with synthetic data...

Area of Science:

• Biomedical Engineering
• Medical Informatics
• Signal Processing

Background:

• Extracting beat-by-beat information from electrocardiograms (ECGs) is vital for diagnostics but current methods struggle with diverse patterns and require extensive annotated data.
• Traditional signal processing and supervised machine learning (ML) methods for ECG delineation face limitations in generalization and data requirements.

Purpose of the Study:

• To develop a robust and generalizable method for ECG detection and delineation.
• To overcome the limitations of existing ECG analysis techniques by leveraging synthetic data generation and novel loss functions.

Main Methods:

• A synthetic data generation scheme that constructs realistic ECG traces from fundamental segments.
• Two novel segmentation-based loss functions designed to improve prediction accuracy and boundary delineation.
• Training a deep learning model using the generated synthetic data and proposed loss functions.

Main Results:

• Achieved a 99.38% F1-score with minimal delineation errors (X ms for onsets, Y ms for offsets) across P, QRS, and T waves.
• Outperformed state-of-the-art methods on three diverse ECG databases (QT, LU, Zhejiang).
• Demonstrated robust generalization across varying lead configurations, sampling frequencies, and pathologies.

Conclusions:

• The proposed approach significantly enhances ECG delineation accuracy and robustness.
• Synthetic data generation and novel loss functions offer a promising solution for data-scarce scenarios in ECG analysis.
• The open-source release of the code facilitates broader application and advancement in clinical ECG analysis.