562 Assessing Intracranial Aneurysm Rupture Risk Using Convolutional Neural Networks Based on Morphological Features

INTRODUCTION: Intracranial aneurysm rupture carries significant mortality and morbidity. This critical neurologic emergency underscores the need for accurate predictive tools to intervene prior to rupture. Current rupture predictions provide annual rupture risk and compare age and size-based rupture risk to procedural risk. METHODS: We developed a convolutional neural network (CNN) to predict aneurysm rupture risk using the Aneux morphology database integrating clinical, morphometric, and hemodynamic data. Clinical features (e.g., age, sex) were one-hot encoded; morphometric features were derived from three-dimensional rotational angiography and reduced via principal components analysis (PCA), with 95% of variance retained; hemodynamic features (e.g., wall shear stress, WSS; oscillatory shear index, OSI) were median-imputed. The dataset was filtered for'dome' cutType. The CNN was trained to classify rupture status, leveraging processed features to enhance prediction accuracy. The model included a dropout of 50%, seven- layers, and a sigmoid output layer. RESULTS: The final dataset included 735 aneurysms (261 ruptured, 474 unruptured) after filtering and deduplication. PCA retained 64 components, explaining 95% variance. The CNN achieved an area under the curve (AUC) of .83, sensitivity of .83, and accuracy of .73. Feature importance was extracted to identify the most relevant morphological features for characterizing rupture status. CONCLUSIONS: The use of CNN approaches for predicting aneurysm rupture risk may have future implications in clinical decision-making by improving risk stratification for interventions. This study adds to aneurysm risk prediction by incorporating morphological data, which has been absent from prior work (e.g., the international Study of Unruptured intracranial Aneurysms), offering a potentially more nuanced understanding of structural contributors to rupture risk.