Abstract Body (Do not enter title and authors here): Background
Silent (subclinical) atrial fibrillation (AF) is common and markedly raises stroke risk, yet it often evades standard screening that relies on symptomatic or incidental detection. We developed and evaluated a deep-learning electrocardiogram (ECG) model that identifies patients harboring paroxysmal AF while in normal sinus rhythm.

Methods
A retrospective cohort of 52,300 12-lead ECGs (10-second, 500 Hz) from 18,240 adults was assembled from two health-system repositories (2015-2023). ECGs acquired within 30 days before or after any documented AF episode were labeled “AF-positive”; all others served as controls after excluding structural heart disease and pacemaker rhythms. A ResNet-34 convolutional neural network was trained on raw voltage–time traces using five-fold cross-validation. The primary end point was area under the receiver-operator curve (AUC) for detecting AF-positive status on an apparently normal ECG. Secondary metrics included sensitivity, specificity, and F1 score at the Youden-optimized threshold.

Results
The model achieved an AUC of 0.83 (95 % CI 0.82 – 0.85), sensitivity 0.78, specificity 0.77, and F1 0.77. In a simulated screening workflow applied to 5,100 primary-care encounters, a single opportunistic ECG flagged 23 % of patients as high-risk, among whom 41 % were clinically diagnosed with AF within 12 months—yielding a positive-predictive value of 0.41 and number-needed-to-screen of 12 to detect one new AF case. Integration into routine ECG interpretation software would add <0.5 seconds processing time and could trigger automatic alerts in primary-care clinics or remote patient-monitoring programs, facilitating targeted rhythm surveillance or anticoagulation evaluation.

Conclusion
A deep-learning ECG model can uncover latent paroxysmal AF from standard sinus-rhythm tracings with strong discriminative performance and practical workflow compatibility. External, multicenter validation is warranted to confirm generalizability and assess impact on downstream stroke-prevention strategies.