Comprehensive Summary
Atrial fibrillation is often silent until it causes stroke or heart failure; therefore, tools that predict who will develop AF could improve targeted monitoring and prevention. Jabbour and their colleagues trained a ResNet model on 669,782 ECGs from 145,323 patients and tested it on a test set of 29,065 patients with 135,544 ECGs. The AUC was 0.78 (95% CI 0.768–0.783). Using a classification threshold of 12%, the model had a sensitivity of 66% and a specificity of 75%. In the test set, 4,360 of 29,065 patients had incident AF within five years. The model flagged about 26% of patients as high risk for incident AF. In a 2,301-subgroup, this ECG AI outperformed current measures, CHARGE AF, in clinical and polygenic scores. External validation in MIMIC IV with 109,870 patients showed similar discrimination with an AUC of 0.77 and a high-risk group hazard ratio of 4.6 (95% CI 4.45–4.74).
Outcomes and Implications
This ECG AI can improve monitoring on a smaller, higher-risk group rather than screening everyone. Clinicians can expect fewer missed cases, but many false positives because of the modest positive predictive value. Clinicians should use the model to prioritize intensified monitoring while not using the output alone to start anticoagulation. With an AUC of 0.78, the tool seems suitable for triage but not for automatic treatment decisions. Prospective trials and cost-effectiveness analyses are needed to show that AI-guided monitoring increases AF detection and reduces stroke.