The electrocardiogram (ECG) is a widely accessible diagnostic tool used to evaluate the electrical activity of the heart. Despite its essential role in diagnosing and monitoring cardiac abnormalities, many healthcare settings still rely on paper-based ECGs. While ECGs are performed at every healthcare level, their transmission throughout the patient continuum is hindered by the diverse range of device manufacturers. Therefore, many settings depend on the transmission of paper copies, which are susceptible to quality deterioration, including ink fading, blurring, folding and stains. Furthermore, ECG images may contain patient-identifiable information and do not meet security and compliance standards. In contrast, digital ECG files can be stored indefinitely while retaining high quality, enabling advanced ECG analysis through modern artificial intelligence algorithms [[1], [2], [3], [4], [5], [6]] and easy, compliant transmission between healthcare professionals.

Several approaches for converting ECG images into digital waveforms have been proposed over the past decades. Initially, ECGScan [7] offered extensive image file compatibility, grid detection, and waveform reconstruction but was limited by significant user input. The fully automated ECG digitization solution from Imperial College London [8] utilized deep learning techniques to convert paper ECGs into digital signals. However, it struggles with lead label detection and has poor performance in low-resolution images or overlapping signals, which may affect its applicability across different ECG abnormalities. Recently, ECGMiner [9], an open-source solution, showed high accuracy in the automated digitization of ECG images. However, it faced challenges such as interpreting lead labels, had issues with overlapping leads, and depended on high-quality image inputs. In summary, most existing approaches require manual user input, are limited to high-quality scans, and fail to cover the complex, real-world spectrum of ECG layouts, lighting conditions, lead visibility, or overlapping waveforms (Table 1) [[7], [8], [9], [10]].

At present, there is no method that can reliably and automatically generate a digital waveform from any 12‑lead ECG image while maintaining robustness under real-world conditions, including variations in input quality, ECG manufacturer, and capture environment. We aimed to develop a versatile, fully automated end-to-end ECG digitization solution deployable via smartphones. We evaluated our approach using a large, independent database of ECG images, ensuring that the evaluation dataset remained entirely separate from those used during model development.