Extracranial vascular tortuosity is considered one of the most relevant factors leading to failed mechanical thrombectomy during stroke treatment. Currently, no objective method exists to reliably identify patients that will present difficulties or impossibility for cervical catheter access during endovascular interventions. Vascular tortuosity may be relevant at different scales, from general descriptors of a complete arterial pathway to local abnormalities which may introduce obstacles impossible to overcome during endovascular interventions. Arterial centerline maps have been widely used to characterize vascular tortuosity, and can be trivially represented as graphs. Graph neural networks offer unique properties that make them a great fit when dealing with these vascular descriptors. In this work, we present ArterialGNet, a graph neural network designed to integrate graph embeddings at multiple scales derived from vascular centerline pathways automatically extracted from CTA. A retrospective dataset comprised of 493 interventions with available CTA, including 19 (3.9%) where cervical access was impossible through transfemoral approach, was used for this study. Our model presents excellent discrimination ability (AUROC = 0.89, 95%CI, 0.88–0.90), outperforming previous approaches. Effective prediction of impossible femoral access may provide decision support to neurointerventionalists, leading to reduced procedural times and improved clinical outcomes in specific patients. Source code is available at https://github.com/perecanals/arterial_gnet.git .