CarotidMamba, a deep learning system analyzing CT angiography scans, achieved 84% accuracy in identifying symptomatic carotid plaques across 689 patients from four medical centers. The AI model demonstrated AUC scores ranging from 0.762 to 0.897 in external validation, consistently outperforming traditional clinical assessments and other AI approaches by significant margins. This represents a meaningful advance in stroke prevention, as current treatment decisions rely heavily on measuring how much an artery is narrowed rather than assessing actual plaque vulnerability. Vulnerable plaques can rupture and cause strokes even when stenosis appears moderate, making this distinction clinically crucial. The multi-center validation across diverse patient populations strengthens the model's potential generalizability, though the retrospective design limits causal inference about clinical outcomes. As a preprint awaiting peer review, these promising results require independent validation before clinical implementation. If confirmed, this technology could enable more precise risk stratification for the 700,000 Americans who suffer strokes annually, potentially identifying high-risk patients who would benefit from intervention despite having less severe stenosis measurements.