A Foundation Transformer Model with Self-Supervised Learning for ECG-Based Assessment of Cardiac and Coronary Function

Le résumé fourni par la source

Background: The wide availability of labeled electrocardiogram (ECG) data has driven major advances in artificial intelligence (AI)-based detection of structural and functional cardiac abnormalities and thus ECG-based diagnosis. However, many critical, high value clinical diagnostic applications, such as assessing myocardial ischemia and coronary microvascular dysfunction, remain underserved due to the limited availability of labeled datasets. We developed a self-supervised ECG foundation model and demonstrate how this approach can overcome this limitation. Methods: A modified vision transformer model was pretrained using a large database of unlabeled ECG waveforms (MIMIC-IV-ECG, N=800,035). The model was then fine-tuned using smaller databases that included high-quality labels derived from positron emission tomography (N=3,126) and clinical reports (N=13,704) for 12 clinical, demographic, and traditional ECG prediction tasks. Diagnostic accuracy and model generalizability were evaluated across five additional cohorts including the publicly available PTB-XL and UK Biobank databases and labels from cardiac magnetic resonance imaging (MRI) and single photon emission computed tomography (SPECT). Results: supervised training. The model retained strong performance across three external and two internal cross-modality databases, with AUROC ranging from 0.771 for impaired MFR to 0.949 for impaired LVEF. Conclusion: This versatile ECG foundation model demonstrates that SSL pretraining enhances diagnostic accuracy and generalizability across diverse cardiac diagnostic applications. By enabling effective learning from limited labeled data, this approach supports AI development for complex but clinically critical tasks, such as detecting myocardial ischemia and coronary microvascular dysfunction, where high-quality labels are costly and scarce.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.