<b>SSMCH-ECG</b>: A Validated Dataset of 12-Lead Paper-Based ECG Images for Cardiac Abnormality Detection

AbstractThe SSMCH-ECG dataset is a clinically validated collection of 849 12-lead paper-based ECG images designed to bridge the gap between traditional paper-based diagnostics and modern deep learning applications. Collected from Shaheed Suhrawardy Medical College and Hospital (SSMCH) in Dhaka, Bangladesh, the dataset comprises three distinct clinical classes: Normal (434), Abnormal (344), and Myocardial Infarction (MI) (71). The ground-truth labels were established in collaboration with a clinical specialist, ensuring high reliability in cardiac abnormality detection. This dataset serves as a vital resource for researchers developing algorithms to digitize and classify paper-based ECGs, particularly in resource-limited settings where digital ECG infrastructure may be lacking. BackgroundElectrocardiograms (ECGs) remain the gold standard for diagnosing cardiovascular diseases (CVDs). While digital ECG acquisition is standard in high-income regions, many healthcare facilities in developing nations still rely on paper-based reports [1]. This creates a "data silo" where critical clinical information is inaccessible to automated diagnostic tools [2]. Recent advancements in Computer Vision and Deep Learning have shown promise in classifying ECGs from images; however, the lack of diverse, high-quality, and clinically validated paper-based datasets limits the generalizability of these models. The SSMCH-ECG dataset addresses this by providing a diverse set of real-world clinical samples [3]. By focusing on multi-class detection (Normal, Abnormal, and MI), this data supports the development of robust models for automated screening, potentially reducing clinicians' workload and improving diagnostic accuracy in primary care settings. MethodsStudy Design and Data Source The SSMCH-ECG dataset was developed as a clinically validated collection of paper-based 12-lead electrocardiogram (ECG) images to support computer vision–based cardiac diagnosis. The dataset was prepared for public dissemination through PhysioNet, ensuring standardized documentation, structured metadata, and reproducibility. Data were collected from Shaheed Suhrawardy Medical College and Hospital. Only routine clinical ECG reports generated during standard diagnostic procedures were considered. No additional interventions or alterations to the clinical workflow were performed. Image Acquisition Paper-based ECG reports were digitized using high-resolution scanning or controlled photography. Acquisition conditions were standardized to maintain: Uniform illuminationMinimal geometric distortionFull visibility of all 12 leadsClear calibration markersAll images were stored in lossless PNG format to preserve signal fidelity.Quality Control A strict exclusion protocol was applied to ensure dataset reliability. ECG images were removed if they contained: Printing artifactsExcessive background noiseCropped or incomplete lead segmentsSevere motion blur or distortionThis quality assurance process minimized bias and ensured suitability for deep learning and signal digitization tasks. Clinical Annotation Ground truth labels were established through manual review by a certified physician using standard cardiological diagnostic criteria. Each ECG was categorized into one of three classes: NormalAbnormalMyocardial Infarction (MI)Data DescriptionThe SSMCH-ECG dataset is a clinically validated collection of 849 high-resolution paper-based 12-lead electrocardiogram (ECG) images acquired from Shaheed Suhrawardy Medical College and Hospital. The dataset was designed to support research in computer vision–based ECG classification, signal digitization, and automated cardiac screening, particularly in resource-constrained healthcare settings. Dataset Composition The dataset contains a total of 849 ECG images categorized into three clinically defined classes: Normal (n = 434): ECGs without significant rhythmic, conduction, or morphological abnormalities.Abnormal (n = 344): ECGs showing arrhythmias, conduction blocks, hypertrophy patterns, or other non-MI pathological deviations.Myocardial Infarction (MI) (n = 71): ECGs demonstrating features consistent with acute or prior myocardial infarction based on established clinical criteria.This multi-class distribution enables the development and evaluation of classification models across varying degrees of cardiac pathology. Image Characteristics Format: JPG (lossless compression)Lead Configuration: Standard 12-lead ECG layoutResolution: High-resolution scans/photographs preserving waveform morphology and calibration markersColor Profile: Original clinical print format (including gridlines and annotations)Each image retains the full diagnostic layout, including waveform traces, calibration signals, and, where available, printed clinical parameters. Usage NotesThe SSMCH-ECG dataset is intended for academic and research use, particularly for developing deep learning and computer vision models for ECG image classification and cardiac abnormality detection. It supports applications such as multi-class classification (Normal, Abnormal, MI), ECG signal digitization from paper-based images, and preprocessing or lead-segmentation benchmarking. Researchers should consider the moderate class imbalance (notably fewer MI samples) when designing experiments and apply appropriate validation strategies. All data are anonymized and must not be used to re-identify patients. EthicsThe research protocol and SSMCH-ECG dataset collection were formally approved by the Research Ethics Committee (REC) of the Faculty of Science and Information Technology at Daffodil International University, Dhaka, Bangladesh (IRB No: REC-FSIT-2025/11474). The study was conducted under the supervision of the Principal Investigator, Md. Hasan Imam Bijoy adhered to strict ethical standards, including obtaining informed consent and ensuring complete anonymity of all paper-based ECG records. Conflicts of InterestThe authors declare that there are no conflicts of interest regarding the publication of this dataset. ReferencesSathi, T. A., Jany, R., Ela, R. Z., Azad, A. K. M., Alyami, S. A., Hossain, M. A., & Hussain, I. (2024). An interpretable electrocardiogram-based model for predicting arrhythmia and ischemia in cardiovascular disease. Results in Engineering, 24, 103381.Kilimci, Z. H., Yalcin, M., Kucukmanisa, A., & Mishra, A. K. (2025). Advancing heart disease diagnosis with vision-based transformer architectures applied to ECG imagery. Image and Vision Computing, 105666.Fatema, K., Montaha, S., Rony, M. A. H., Azam, S., Hasan, M. Z., & Jonkman, M. (2022). A robust framework combining image processing and deep learning hybrid model to classify cardiovascular diseases using a limited number of paper-based complex ECG images. Biomedicines, 10(11), 2835.