ECGAug: A novel method of generating augmented annotated electrocardiogram QRST complexes and rhythm strips

Area of Science:

• Computational Biology and Bioinformatics
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Neural networks (NNs) show increasing medical applications, particularly in electrocardiogram (ECG) analysis.
• Manual annotation of ECG fiducial points for NN training is time-consuming and often prohibitive.
• Current time series data augmentation methods risk generating non-physiological signals and lack accurate fiducial point localization.

Purpose of the Study:

• To develop a novel method, ECGAug, for generating augmented ECG training data.
• To create physiologically plausible QRST signals with accurate fiducial point annotations.
• To improve the performance of NNs used for ECG segmentation and annotation.

Main Methods:

• ECGAug recombines existing annotated QRS complexes and T waves from a library.
• The algorithm applies physiological transformations to generate new, annotated QRST complexes and rhythm strips.
• The generated data is used to augment training datasets for NN models.

Main Results:

• ECGAug successfully generated a dataset of physiologically realistic, annotated QRST signals.
• Training NNs with ECGAug-augmented data significantly improved QRST complex annotation performance compared to limited original datasets.
• The efficacy of ECGAug was demonstrated in experiments, showing enhanced NN performance.

Conclusions:

• ECGAug provides an effective solution for generating augmented ECG data with accurate fiducial point annotations.
• The method addresses limitations of existing augmentation techniques, enabling better NN training for ECG analysis.
• ECGAug software is open-source, facilitating its adoption for improving ECG annotation algorithms.