What does this research mean for the field?

A large-scale dataset of paired ECG images containing diverse physical and imaging artifacts alongside ground-truth digital time series enables robust training and benchmarking of automated ECG digitization and classification methods. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To introduce ECG-Image-Database, a public resource that couples electrocardiogram (ECG) images with digital time-series signals to overcome real-world physical and scanning artifacts during automated digitization.

July 2, 2026Physiological MeasurementOpen Access

ECG-image-database: large-scale paired ECG images and time-series with real-world artifacts; a foundation for computerized ECG digitization and analysis

View Full Paper

Ask AI

Bookmark

Why the study?

A database reflecting real-world digitization challenges with diverse scanning, imaging, and physical artifacts paired with time-series signals was needed.

Key Points

To introduce ECG-Image-Database, a public resource that couples electrocardiogram (ECG) images with digital time-series signals to overcome real-world physical and scanning artifacts during automated digitization.
Converted 1,977 digital 12-lead ECG time-series records (PTB-XL, Germany, N=977; Emory Healthcare, USA, N=1,000) into realistic printouts using ECG-Image-Kit, applying synthetic distortions (noise, wrinkles) and physical degradations (soaking, staining, mold, screen photos) to yield 35,595 images.
Acquired 1,596 photographed and scanned images from natural thermal-paper ECG printouts obtained from 266 patients during routine care at Akershus University Hospital in Norway.
The compiled ECG-Image-Database features 37,191 images generated from 2,243 ECG records across three countries, encompassing diverse layouts, lighting, and degradation levels.
Each image is matched with its underlying digital time series to support supervised model benchmarking, serving as the benchmark foundation for the PhysioNet Challenge 2024 and Kaggle 2025–2026 competitions.

Structured PICO

Population

2,243 ECG records from three sources: the PTB-XL dataset (n=977, Germany), Emory Healthcare (n=1,000, USA), and Akershus University Hospital (n=266, Norway).

Intervention

Development of ECG-Image-Database using ECG-Image-Kit to generate paired ECG images and underlying time-series signals with controlled digital and physical artifacts.

Outcome

Creation of a large-scale database comprising 37,191 images derived from 2,243 ECG records across varying artifact conditions and capture modalities.

The creation of the ECG-Image-Database provides a critical large-scale, open-source resource of real-world distorted ECG images paired with exact time-series data to train and validate machine learning models for ECG digitization.

Implication

May support AI ECG digitization development; leaves open real-world clinical impact and accuracy gains.

Abstract

Objective. We present electrocardiogram (ECG)-Image-Database, comprising ECG image datasets paired with the underlying time-series signals. The database is designed to reflect real-world digitization challenges by incorporating diverse scanning, imaging, and physical artifacts commonly encountered in clinical and archival ECG records. It currently includes ECG images of records from Emory Healthcare (Atlanta, USA), the PTB-XL dataset from Physikalisch-Technische Bundesanstalt (Berlin, Germany), and Akershus University Hospital (Lørenskog, Norway). Approach. ECG images from Emory Healthcare and PTB-XL were generated using ECG-Image-Kit, an open-source toolkit that converts 12-lead ECG time series into realistic printout-style images. The toolkit introduces controlled digital distortions, including noise, wrinkles, stains, and perspective changes. A total of 1977 ECG records (977 PTB-XL and 1000 Emory Healthcare) were processed to create high-fidelity ECG printouts. These images were further subjected to physical degradation—such as soaking, staining, and mold exposure—before being scanned or photographed under varying lighting and capture conditions. In addition, we captured mobile photographs of the original ECG images displayed on computer monitors. This process resulted in 35 595 images. Independently, ECG images from Akershus University Hospital were captured from thermal-paper ECG printouts from 266 patients acquired during routine clinical care. Each printout underwent natural handling degradation and was repeatedly scanned and photographed, yielding 1596 images. Main results. ECG-Image-Database contains 37 191 images derived from 2243 ECG records across three countries. The dataset spans clean scans, noisy photographs, and severe physical deterioration, and includes heterogeneous ECG layouts and acquisition conditions. Each record is linked to its corresponding digital ECG time series, enabling supervised training and benchmarking of digitization and classification methods. Subsets of this resource were used in the PhysioNet Challenge 2024 and a subsequent 2025–2026 Kaggle challenge on ECG image digitization and classification. The database is structured for scalability and future expansion. Significance. By providing paired ECG images and signals across diverse real-world artifact conditions, ECG-Image-Database supports robust development of methods for recovering ECG time series from images and for direct image-based interpretation, helping preserve diagnostic value from non-digital ECG archives.