Abstract Background and aims Size-based heuristics and the absence of patient-specific flow characterization limit rupture risk assessment for intracranial aneurysms (IA). Many flow patterns linked to rupture traditionally require invasive angiography. This study tests whether integrated features from non-invasive hemodynamics, morphology, and clinical history can capture these signatures and improve prediction. Methods 71 IA patients with saccular aneurysms were retrospectively analyzed (n=71; 10 ruptured, 61 unruptured, n aneurysms). Expert segmentation enabled patient-specific vascular models for non-invasive CFD simulations, yielding hemodynamic (TAWSS, low-WSS area%, OSI, RRT) and morphological features (aspect ratio, size ratio, irregularity, volume, neck and dome geometry). Limited PHI (age, sex, hypertension, prior SAH) was extracted from chart reviews. Features were integrated into a composite Risk-of-Rupture (RoR) score (1-10) via ensemble machine learning (Logistic Regression, Random Forest, Gradient Boosting) with cross-validation. Results RoR achieved AUC 0.93 (88% sensitivity, 85% specificity), significantly outperforming PHASES (AUC 0.66) and UIATS (AUC 0.62). Risk stratification classified aneurysms as Low-risk (64%), Moderate-risk (34%), and High-risk (2%), enabling precise triage. Model retained strong performance (AUC 0.93) under minimal PHI and demonstrated excellent calibration across age, morphology, and clinical subgroups. Conclusions Ruptured aneurysms exhibited characteristic geometric risk factors and hemodynamic profiles (depressed WSS, elevated OSI/RRT) identified non-invasively via CFD. Among unruptured cases, integrated profiling revealed concealed moderate-risk phenotypes missed by morphology alone. These findings demonstrate that imaging-derived non-invasive analysis delivers physiological insights comparable to invasive angiography, enabling precise risk stratification to guide surveillance and treatment decisions. Conflict of interest Prithvinath Reddy Garigapuram: nothing to disclose
Garigapuram et al. (2026) studied this question.