Machine learning models incorporating longitudinal patient data achieved superior performance predicting major cardiovascular events in atrial fibrillation patients compared to established clinical scoring systems. The AI approach reached 65% accuracy (AUC 0.65) for one-year stroke prediction versus 59% for the widely-used CHA2DS2-VASc score, and 78% accuracy for mortality prediction against 72% for GARFIELD-AF reference scores. The models analyzed six clinical endpoints including stroke, death, heart failure hospitalizations, and acute coronary syndrome in a Portuguese cohort. This represents meaningful progress in cardiovascular risk assessment, as atrial fibrillation affects millions globally and current prediction tools often miss the dynamic nature of individual risk evolution. The longitudinal approach captures how risk factors change over time rather than relying on static snapshots. However, the modest improvement over existing tools raises questions about clinical utility versus implementation complexity. The study's focus on a single population limits generalizability, and as an unreviewed preprint, these promising results require validation through peer review and broader cohort testing before clinical adoption.