Atrial fibrillation (AF) is a common cause of ischemic stroke, accounting for up to one-third of all cases. Untreated AF can increase the risk of stroke by up to five times and make stroke recurrence more likely. Anticoagulation has proven beneficial in reducing stroke risk. However, AF is often paroxysmal and asymptomatic, remaining undetected and undiagnosed in up to 30% of cases. The current methods for AF detection are usually lengthy (cardiac monitoring), expensive (smart devices), or invasive (implantable cardiac monitors), limiting their routine use. We present a novel method to screen for AF by analyzing infarct patterns of stroke patients from brain magnetic resonance imaging (MRI) scans. We propose EDAF, a novel method based on the segment anything model (SAM) that leverages the power of a foundational deep learning model to efficiently analyze brain MRI and identify whether the underlying stroke etiology is AF. EDAF is trained and validated using a retrospectively acquired dataset of 235 post-stroke patients, achieving an area under the receiver operating characteristic (AUROC) of \(83.08\% \pm 2.96\% \) in identifying the presence of AF. EDAF can achieve optimal solutions with minimal training, highlighting its potential for use in low-resource settings. As MRI is readily available in stroke centers and routinely performed on many patients after a stroke, either during their admission or as an outpatient, the proposed method can effectively identify patients for further AF investigation.