Related Experiment Video
Updated: May 12, 2026
08:22
BrainBeats as an Open-Source EEGLAB Plugin to Jointly Analyze EEG and Cardiovascular Signals
Published on: April 26, 2024
SynthECG: Python Framework and ECG Image Datasets for Digitization, Lead Detection, and Waveform Segmentation
Masoud Rahimi1, Reza Karbasi1, Abdol-Hossein Vahabie1
1Department of Machine Intelligence and Robotics, School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran.
Journal of Medical Signals and Sensors
|May 11, 2026
Summary
SynthECG, an open-source framework, generates synthetic electrocardiogram (ECG) datasets for deep learning tasks like digitization and lead detection. This tool addresses challenges with realistic overlapping waveforms, providing valuable resources for AI in healthcare.
Area of Science:
- Biomedical Engineering
- Artificial Intelligence in Medicine
- Medical Imaging Analysis
Background:
- Digitizing electrocardiogram (ECG) images into structured time-series data is crucial for clinical applications.
- Existing datasets lack standardization, particularly for complex scenarios like overlapping waveforms, hindering deep learning model development.
- Accurate ECG analysis requires robust methods for signal extraction and interpretation from visual representations.
Purpose of the Study:
- To introduce SynthECG, an open-source Python framework for generating synthetic ECG image datasets.
- To create tailored datasets for deep learning tasks including ECG digitization, lead detection, and waveform segmentation.
- To address the challenge of overlapping waveforms in ECG images through a novel simulation mechanism.
Main Methods:
- Developed SynthECG, a customizable framework supporting various ECG dataset generation parameters.
- Implemented deep learning-ready dataset generation for ECG digitization, lead/lead name detection (YOLO), and waveform segmentation (U-Net).
- Introduced a novel method to simulate overlapping waveforms from adjacent leads while maintaining segmentation mask integrity.
Main Results:
- Generated four open-access datasets: ECG digitization (2000 images), lead/name detection (2000 images), normal segmentation (20,000 images), and overlapping waveforms (102 images).
- Validated the digitization dataset with a non-ML algorithm, achieving a mean squared error of 0.002 and high correlation (ρ: 0.93).
- Demonstrated the utility of generated datasets for lead/name detection using YOLOv8.
Conclusions:
- SynthECG provides a scalable solution for generating customizable ECG image datasets for deep learning.
- The framework supports critical tasks like digitization (normal and overlapping conditions) and lead detection.
- Publicly available datasets and code facilitate research in AI-driven ECG analysis.
