Autoregressive prediction of hemodynamics before TAVR in latent space with pretrained video autoencoders

Transcatheter aortic valve replacement (TAVR) is increasingly applied to intermediate- and low-risk patients, raising the need for better prediction of long-term complications such as aortopathy and thrombosis, as well as efficient, patient-specific hemodynamic assessment. At the same time, recent advances in large video generative models demonstrate that videos can be compressed into low-dimensional latent spaces that preserve key spatiotemporal structure. We hypothesize that pretrained autoencoders can compress cardiovascular blood-flow dynamics and enable prediction of their temporal evolution in latent space. We evaluate variational autoencoders (VAEs) from state-of-the-art video models and train an autoregressive 3D U-Net on the latent space to predict systolic flow in a synthetic virtual cohort of 100 TAVR patients. The patients were generated from statistical shape and calcification models, and unsteady systolic flow was simulated by means of computational fluid dynamics (CFD). The resulting velocity fields were voxelized and slice-wise encoded with the VAEs. On the latent space 3D U-Nets were trained to autoregressively predict future latent time steps, which were decoded back to velocity fields and compared with CFD simulations. Our results show that pretrained VAEs can compress systolic hemodynamics while preserving relevant flow patterns, thereby enabling forecasting and suggesting a promising route toward computationally efficient surrogates for TAVR hemodynamic assessment.