Identification of anatomical and angio-architectural features of at-risk cranial dural arteriovenous fistulas using machine learning approaches

BackgroundCranial dural arteriovenous fistulas (dAVFs) are rare complex vascular malformations that have a bleeding risk with potential lethal consequences. Despite this, the vascular architectural features associated with the rupture risk are not always clearly defined. MethodsWe retrospectively analyzed cranial arteriovenous fistulas in terms of their anatomical and angio-architectural features as evaluated on conventional subtraction angiography: Location of the fistula, fistula architecture, venous ectasia, reflux in cortical draining veins, presence of pial feeders, outflow stenosis, presence of a major sinus thrombosis, flow-associated arterial aneurysms as well as presenting symptoms. Patterns in the data were identified after multiple components analysis followed by automatic k-means clustering and their predictive power was confirmed using a neural network and a random forest classifier. ResultsNew relevant features predictive of hemorrhage (venous outflow stenosis and fistula architecture) were identified using distinct but surprisingly converging modeling paradigms. Both the neural network and the random forest classifier achieved a relatively high performance metric, with area under the receiver operating characteristic curve (ROC AUC)) of 0.875 [95% CI, 0.75-1.0]. The relevance of these findings was verified by performing a multiple correspondence analysis followed by k-means clustering in the angiographic feature vector space. There was good agreement between the ground truth (hemorrhage) and the cluster labels (adjusted Rand score 0.273, purity index 0.82). ConclusionMachine learning approaches confirmed the importance of previously described features (reflux in a cortical vein and venous ectasia) but also uncovered novel relevant characters (outflow stenosis and fistula architecture) for the hemorrhage risk of dAVFs.