Machine Learning-Based Reduced-Order Prediction of Rupture-Related Hemodynamics in Cerebral Aneurysms with Surface Blebs: A Pilot Study.

Cerebral aneurysm rupture is strongly associated with localized hemodynamic abnormalities, particularly in aneurysms containing surface blebs that represent structurally weakened regions of the vessel wall. Although computational fluid dynamics (CFD) provides detailed insight into aneurysm hemodynamics, its high computational cost limits its application in rapid patient-specific risk assessment. This pilot study presents a POD-LSTM reduced-order framework for the analysis and temporal prediction of CFD-derived hemodynamics in patient-specific cerebral aneurysms with surface blebs. Four patient-specific aneurysm geometries located in the anterior cerebral artery (ACA), middle cerebral artery (MCA), internal carotid artery (ICA), and basilar artery (BAS) were selected from the Aneurisk database. High-fidelity CFD simulations were first performed under pulsatile, laminar, incompressible, non-Newtonian blood flow conditions using the Casson viscosity model. Regions experiencing elevated wall shear stress (WSS) were identified, and surface blebs were subsequently introduced to evaluate their influence on intra-aneurysmal flow characteristics. Proper orthogonal decomposition (POD) was employed to reduce the dimensionality of the transient CFD datasets, while a long short-term memory (LSTM) neural network was trained to predict the temporal evolution of the dominant POD coefficients and reconstruct the hemodynamic fields. The POD analysis demonstrated that more than 99% of the flow energy was captured using fewer than 20 modes for pressure and WSS and approximately 25-30 modes for the remaining variables. The POD-LSTM framework accurately reconstructed velocity, pressure, WSS, oscillatory shear index (OSI), vorticity, and helicity with consistently low reconstruction and prediction errors. Comparisons with CFD results showed excellent agreement in both temporal evolution and spatial distributions of the hemodynamic parameters. Hemodynamic analyses further revealed that bleb formation generally increased local wall shear loading and neck velocity while altering the distributions of OSI, relative residence time (RRT), and vorticity, producing localized regions of elevated mechanical stress associated with increased rupture susceptibility. The proposed POD-LSTM framework substantially reduces computational cost while maintaining high prediction accuracy, providing a compact reduced-order framework for efficient representation and temporal prediction of CFD-derived aneurysm hemodynamics. However, the limited number of cases and absence of clinical rupture outcomes preclude statistical generalization and direct clinical risk prediction.