AI-Driven Intracranial Aneurysm Rupture Risk Prediction through Morphology and Hemodynamics

Introduction/Purpose Intracranial aneurysms affect 3‐5% of the population and account for approximately 30,000ruptures annually in the U.S., with a 50% 30‐day mortality rate and permanent disability in most survivors. Current clinical decision support tools (CDSTs), including PHASES and UIATS, rely on static, population‐level factors and may misclassify high‐risk aneurysms. These models fail to incorporate longitudinal morphological evolution or patient‐specific hemodynamics. We aim to develop and validate an AI‐driven Risk of Rupture (RoR) module that integrates clinical, imaging, and physiologic features to improve rupture prediction. Materials/Methods This retrospective study includes 1,000 patients (500 ruptured, 500 unruptured) with intracranial aneurysms and up to 5 years of imaging and clinical follow‐up. Inputs include demographics, medical history, morphologic features (size, aspect ratio, irregularity), and computational fluid dynamics (CFD)‐derived hemodynamic indices (velocity, wall shear stress, oscillatory shear index, total pressure). Imaging will be processed with an AI‐assisted segmentation pipeline that reduces processing time from 45 minutes to 15 seconds. Longitudinal data will be harmonized into a FHIR‐compliant dataset. A machine learning classifier will be trained using State Space Self‐Organizing Maps and competing risk survival modeling. Model interpretability was assessed using SHAP analysis to identify key rupture predictors. Performance will be benchmarked against PHASES and UIATS using AUC‐ROC and calibration metrics. Results In a pilot dataset of 412 aneurysms (211 ruptured, 201 unruptured), the RoR module achieved91% classification accuracy and an AUC‐ROC of 0.90, significantly outperforming PHASES (61%accuracy, AUC 0.52) and UIATS (63% accuracy, AUC 0.53). The model consistently highlighted aneurysm size, bifurcation location, and dome irregularity as top contributors to rupture risk, aligning with clinical intuition. The segmentation package achieved a Dice coefficient of 0.88 and IoU of 0.80 across modalities. Cross‐validation confirmed statistically significant improvements compared with existing tools ( p < 0.05). Conclusion The RoR module has thus far demonstrated superior rupture risk prediction compared with current CDSTs, leveraging longitudinal multimodal data and explainable AI. This tool has the potential to reduce overtreatment, enable earlier identification of high‐risk aneurysms, and support scalable EHR integration as a clinician‐friendly decision‐support platform.