Longitudinal modeling of neurodegenerative diseases rem- ains a major challenge due to the complexity of the underlying processes and the strong heterogeneity of individual trajectories. Neuroimaging provides a unique opportunity to characterize disease-related alterations long before clinical onset, enabling earlier intervention. In this work, we employ a Variational Autoencoder (VAE) to compress high-dimensional neuroimaging scans into a low-dimensional latent representation ( \(z_0\) ) that captures dementia-related patterns. We then model the temporal evolution of this neuroimaging biomarker using a Bayesian longitudinal framework that incorporates genetic information as covariates. This approach allows us to study population-level disease trajectories while explicitly accounting for genetic risk factors. Our results show that the proposed neuroimaging biomarker \(z_0\) and the Bayesian longitudinal model successfully stratifies subjects into genotype-dependent risk subgroups, revealing distinct baseline profiles and progression patterns consistent with known genetic effects.