Abstract

Introduction:Early detection of cardiovascular disease and accurate risk prediction are critical for improving patient outcomes.While AI-enhanced ECG (AI-ECG) tools show great promise, many depend on digital signal input, limiting their clinical utility in environments lacking.This study introduces an image-based AI-ECG platform for cardiac screening and risk prediction.Methods: We developed deep learning models using over 1.1 million ECG images from 189,538 patients in a secondary care dataset.Model performance was validated internally and externally across multiple cohorts: CODE (n = 645,380), SaMi-Trop (n = 1,022), and UK Biobank (n = 42,386).Robustness was tested using photographed ECGs, lead position variation, simulated artifacts, and incomplete leads.Results: The platform demonstrated strong diagnostic capabilities, including detection of reduced ejection fraction (AUC 0.896) and valvular heart diseases: aortic stenosis (AUC 0.869), tricuspid regurgitation (0.850), mitral regurgitation (0.826), and aortic regurgitation (0.775).Mortality prediction showed strong performance across all cohorts (c-statistic 0.818 [0.814-0.821] in BIDMC, 0.832 [0.814-0.850] in photographed images, 0.764 [0.761-0.768] in CODE and 0.770 [0.725-0.816] in SaMi-Trop).Future cardiac conditions were accurately predicted in internal validation: atrial fibrillation (0.770 [0.764-0.776]),heart failure (0.809 [0.805-0.814]),ventricular arrhythmia (0.800 [0.788-0.814]),atherosclerotic cardiovascular disease (0.758 [0.752-0.765]),and complete heart block (0.855 [0.840-0.870]).High-risk quartiles consistently showed significantly elevated event risks.Saliency mapping revealed interpretable and condition-specific activation patterns.Conclusion: This image-based AI-ECG platform offers robust, scalable, and interpretable cardiac risk prediction using standard ECG images.Its generalisability across diverse cohorts and image types makes it suitable for global clinical deployment, particularly in low-resource settings.