Diagnostic Accuracy of Vision Transformer AI in Assessing Myocardial Flow Reserve

Article Impact Level: HIGH
Data Quality: STRONG
Summary of  NEJM AI  https://doi.org/10.1056/AIoa2500164 
Dr. Jonathan B. Moody  et al.

Points

Researchers utilized self-supervised learning on a massive unlabeled dataset to train a vision transformer model for detecting complex cardiac conditions via standard electrocardiograms.
The model demonstrated an area under the receiver operator curve of 0.763 for identifying impaired myocardial flow reserve and 0.955 for reduced left ventricular ejection fraction.
Diagnostic accuracy improved significantly in eleven out of twelve clinical prediction tasks when compared to conventional de novo supervised training methods used in previous studies.
This tool effectively identifies coronary microvascular dysfunction using a ten-second strip which offers a rapid alternative to expensive and often inaccessible positron emission tomography scans.
The foundation model retained strong diagnostic performance across multiple external databases including the UK Biobank and PTB-XL to validate its generalizability in diverse clinical settings.

Summary

This study evaluated a self-supervised ECG foundation model designed to detect coronary microvascular dysfunction (CMVD) and other functional abnormalities often missed by standard diagnostics. Addressing the scarcity of labeled data from advanced imaging, researchers pretrained a modified vision transformer on 800,035 unlabeled ECG waveforms before fine-tuning it with labeled data from positron emission tomography (N=3,126) and clinical reports (N=13,704). The objective was to determine if a deep learning model could accurately predict myocardial flow reserve (MFR) and left ventricular ejection fraction (LVEF) using standard 10-second ECG strips.

Diagnostic performance was assessed across 12 clinical tasks, with the model achieving an area under the receiver operator curve (AUROC) of 0.763 for detecting impaired MFR (<2) and 0.955 for impaired LVEF (<35%). Self-supervised learning (SSL) pretraining significantly enhanced diagnostic accuracy in 11 of the 12 tasks compared to conventional supervised training methods. Validation across five external cohorts, including the UK Biobank and PTB-XL databases, demonstrated robust generalizability, with AUROC values ranging from 0.771 for impaired MFR to 0.949 for impaired LVEF.

These results indicate that SSL pretraining enables effective learning from limited high-value labeled datasets to predict complex cardiac conditions. By facilitating the detection of myocardial ischemia and microvascular dysfunction without expensive imaging, this tool offers a cost-effective, noninvasive method for identifying high-risk patients in resource-limited settings. The model consistently outperformed previous state-of-the-art AI tools, establishing a new framework for ECG-based assessment of cardiac function.

Link to the article: https://ai.nejm.org/doi/10.1056/AIoa2500164

References

Moody, J. B., Poitrasson-Rivière, A., Renaud, J. M., Hagio, T., Alahdab, F., Al-Mallah, M. H., Vanderver, M. D., Goonewardena, S. N., Ficaro, E. P., & Murthy, V. L. (2025). A foundation transformer model with self-supervised learning for ecg-based assessment of cardiac and coronary function. NEJM AI, 2(12). https://doi.org/10.1056/AIoa2500164