Abstract Body: Introduction. 3 to 5% of patients undergoing endovascular thrombectomy present impossible catheter access to the occlusion site from transfemoral access (TFA), largely attributed to complex arterial anatomy. Radial access can be an effective bailout strategy, but intraprocedural delays may negatively impact outcomes. Novel image processing algorithms allow for advanced characterization of vascular pathways from baseline neuroimaging, enabling the exploration of predictive models of impossible TFA before arterial puncture.
Methods. A retrospective cohort of patients with an anterior large vessel occlusion who received thrombectomy from TFA between 2017 and 2023 were included in this study. A previously described automatic vascular analysis software was used to generate centerline graphs from the aorta to the intracranial occlusion site from baseline CTA. ArterialGNet, a graph neural network based on graph attention designed to integrate descriptors of centerline pathways extracted at three different distance scales, was trained for impossible TFA prediction. Five-fold cross validation was used for model derivation. The method was compared to a previously introduced random forest ensemble model with extreme gradient boosting (XGBRF) based on six vascular tortuosity descriptors of the aortic and supra-aortic regions.
Results. A total of 745 patients (aged 78 years IQR 68-85, 56% women) were included in this study. Patients treated between 2017 to 2022 (n=568, 3.2% with impossible TFA) were used for model training and validation. Patients treated in 2023 (n=177, 3.4% with impossible TFA) were held out for testing. In validation, the best-performing configuration of ArterialGNet achieved a C-statistic of 0.82 (95%CI 0.74-0.90), similar to the baseline model (0.82, 95%CI 0.77-0.88). Comparable outcomes were observed in the final testing for ArterialGNet (0.84, 95%CI: 0.82–0.86). In contrast, the XGBRF model exhibited signs of overfitting (0.65, 95% CI: 0.53–0.78). In final testing, ArterialGNet predicted impossible TFA with a sensitivity of 0.80 (95%CI 0.66-0.94) and a specificity of 0.84 (95%CI 0.76-0.91). Median processing time for ArterialGNet was below 4 min.
Conclusions. A novel model for impossible TFA prediction was validated with a large dataset. Impossible TFA prediction before arterial puncture may assist in decision support for initial access selection in thrombectomy, reducing intraprocedural delays and potentially improving clinical outcomes.